fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/compat.h>
  34 #include <linux/bit_spinlock.h>
  35 #include <linux/xattr.h>
  36 #include <linux/posix_acl.h>
  37 #include <linux/falloc.h>
  38 #include <linux/slab.h>
  39 #include <linux/ratelimit.h>
  40 #include <linux/mount.h>
  41 #include <linux/btrfs.h>
  42 #include <linux/blkdev.h>
  43 #include <linux/posix_acl_xattr.h>
  44 #include <linux/uio.h>
  45 #include <asm/unaligned.h>
  46 #include "ctree.h"
  47 #include "disk-io.h"
  48 #include "transaction.h"
  49 #include "btrfs_inode.h"
  50 #include "print-tree.h"
  51 #include "ordered-data.h"
  52 #include "xattr.h"
  53 #include "tree-log.h"
  54 #include "volumes.h"
  55 #include "compression.h"
  56 #include "locking.h"
  57 #include "free-space-cache.h"
  58 #include "inode-map.h"
  59 #include "backref.h"
  60 #include "hash.h"
  61 #include "props.h"
  62 #include "qgroup.h"
  63 #include "dedupe.h"
  64
  65 struct btrfs_iget_args {
  66         struct btrfs_key *location;
  67         struct btrfs_root *root;
  68 };
  69
  70 struct btrfs_dio_data {
  71         u64 outstanding_extents;
  72         u64 reserve;
  73         u64 unsubmitted_oe_range_start;
  74         u64 unsubmitted_oe_range_end;
  75         int overwrite;
  76 };
  77
  78 static const struct inode_operations btrfs_dir_inode_operations;
  79 static const struct inode_operations btrfs_symlink_inode_operations;
  80 static const struct inode_operations btrfs_dir_ro_inode_operations;
  81 static const struct inode_operations btrfs_special_inode_operations;
  82 static const struct inode_operations btrfs_file_inode_operations;
  83 static const struct address_space_operations btrfs_aops;
  84 static const struct address_space_operations btrfs_symlink_aops;
  85 static const struct file_operations btrfs_dir_file_operations;
  86 static const struct extent_io_ops btrfs_extent_io_ops;
  87
  88 static struct kmem_cache *btrfs_inode_cachep;
  89 struct kmem_cache *btrfs_trans_handle_cachep;
  90 struct kmem_cache *btrfs_path_cachep;
  91 struct kmem_cache *btrfs_free_space_cachep;
  92
  93 #define S_SHIFT 12
  94 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  95         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  96         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  97         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  98         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  99         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
 100         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 101         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 102 };
 103
 104 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 105 static int btrfs_truncate(struct inode *inode);
 106 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 107 static noinline int cow_file_range(struct inode *inode,
 108                                    struct page *locked_page,
 109                                    u64 start, u64 end, u64 delalloc_end,
 110                                    int *page_started, unsigned long *nr_written,
 111                                    int unlock, struct btrfs_dedupe_hash *hash);
 112 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
 113                                        u64 orig_start, u64 block_start,
 114                                        u64 block_len, u64 orig_block_len,
 115                                        u64 ram_bytes, int compress_type,
 116                                        int type);
 117
 118 static void __endio_write_update_ordered(struct inode *inode,
 119                                          const u64 offset, const u64 bytes,
 120                                          const bool uptodate);
 121
 122 /*
 123  * Cleanup all submitted ordered extents in specified range to handle errors
 124  * from the fill_dellaloc() callback.
 125  *
 126  * NOTE: caller must ensure that when an error happens, it can not call
 127  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 128  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 129  * to be released, which we want to happen only when finishing the ordered
 130  * extent (btrfs_finish_ordered_io()). Also note that the caller of the
 131  * fill_delalloc() callback already does proper cleanup for the first page of
 132  * the range, that is, it invokes the callback writepage_end_io_hook() for the
 133  * range of the first page.
 134  */
 135 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 136                                                  const u64 offset,
 137                                                  const u64 bytes)
 138 {
 139         unsigned long index = offset >> PAGE_SHIFT;
 140         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 141         struct page *page;
 142
 143         while (index <= end_index) {
 144                 page = find_get_page(inode->i_mapping, index);
 145                 index++;
 146                 if (!page)
 147                         continue;
 148                 ClearPagePrivate2(page);
 149                 put_page(page);
 150         }
 151         return __endio_write_update_ordered(inode, offset + PAGE_SIZE,
 152                                             bytes - PAGE_SIZE, false);
 153 }
 154
 155 static int btrfs_dirty_inode(struct inode *inode);
 156
 157 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 158 void btrfs_test_inode_set_ops(struct inode *inode)
 159 {
 160         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 161 }
 162 #endif
 163
 164 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 165                                      struct inode *inode,  struct inode *dir,
 166                                      const struct qstr *qstr)
 167 {
 168         int err;
 169
 170         err = btrfs_init_acl(trans, inode, dir);
 171         if (!err)
 172                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 173         return err;
 174 }
 175
 176 /*
 177  * this does all the hard work for inserting an inline extent into
 178  * the btree.  The caller should have done a btrfs_drop_extents so that
 179  * no overlapping inline items exist in the btree
 180  */
 181 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 182                                 struct btrfs_path *path, int extent_inserted,
 183                                 struct btrfs_root *root, struct inode *inode,
 184                                 u64 start, size_t size, size_t compressed_size,
 185                                 int compress_type,
 186                                 struct page **compressed_pages)
 187 {
 188         struct extent_buffer *leaf;
 189         struct page *page = NULL;
 190         char *kaddr;
 191         unsigned long ptr;
 192         struct btrfs_file_extent_item *ei;
 193         int ret;
 194         size_t cur_size = size;
 195         unsigned long offset;
 196
 197         if (compressed_size && compressed_pages)
 198                 cur_size = compressed_size;
 199
 200         inode_add_bytes(inode, size);
 201
 202         if (!extent_inserted) {
 203                 struct btrfs_key key;
 204                 size_t datasize;
 205
 206                 key.objectid = btrfs_ino(BTRFS_I(inode));
 207                 key.offset = start;
 208                 key.type = BTRFS_EXTENT_DATA_KEY;
 209
 210                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 211                 path->leave_spinning = 1;
 212                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 213                                               datasize);
 214                 if (ret)
 215                         goto fail;
 216         }
 217         leaf = path->nodes[0];
 218         ei = btrfs_item_ptr(leaf, path->slots[0],
 219                             struct btrfs_file_extent_item);
 220         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 221         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 222         btrfs_set_file_extent_encryption(leaf, ei, 0);
 223         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 224         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 225         ptr = btrfs_file_extent_inline_start(ei);
 226
 227         if (compress_type != BTRFS_COMPRESS_NONE) {
 228                 struct page *cpage;
 229                 int i = 0;
 230                 while (compressed_size > 0) {
 231                         cpage = compressed_pages[i];
 232                         cur_size = min_t(unsigned long, compressed_size,
 233                                        PAGE_SIZE);
 234
 235                         kaddr = kmap_atomic(cpage);
 236                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 237                         kunmap_atomic(kaddr);
 238
 239                         i++;
 240                         ptr += cur_size;
 241                         compressed_size -= cur_size;
 242                 }
 243                 btrfs_set_file_extent_compression(leaf, ei,
 244                                                   compress_type);
 245         } else {
 246                 page = find_get_page(inode->i_mapping,
 247                                      start >> PAGE_SHIFT);
 248                 btrfs_set_file_extent_compression(leaf, ei, 0);
 249                 kaddr = kmap_atomic(page);
 250                 offset = start & (PAGE_SIZE - 1);
 251                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 252                 kunmap_atomic(kaddr);
 253                 put_page(page);
 254         }
 255         btrfs_mark_buffer_dirty(leaf);
 256         btrfs_release_path(path);
 257
 258         /*
 259          * we're an inline extent, so nobody can
 260          * extend the file past i_size without locking
 261          * a page we already have locked.
 262          *
 263          * We must do any isize and inode updates
 264          * before we unlock the pages.  Otherwise we
 265          * could end up racing with unlink.
 266          */
 267         BTRFS_I(inode)->disk_i_size = inode->i_size;
 268         ret = btrfs_update_inode(trans, root, inode);
 269
 270 fail:
 271         return ret;
 272 }
 273
 274
 275 /*
 276  * conditionally insert an inline extent into the file.  This
 277  * does the checks required to make sure the data is small enough
 278  * to fit as an inline extent.
 279  */
 280 static noinline int cow_file_range_inline(struct btrfs_root *root,
 281                                           struct inode *inode, u64 start,
 282                                           u64 end, size_t compressed_size,
 283                                           int compress_type,
 284                                           struct page **compressed_pages)
 285 {
 286         struct btrfs_fs_info *fs_info = root->fs_info;
 287         struct btrfs_trans_handle *trans;
 288         u64 isize = i_size_read(inode);
 289         u64 actual_end = min(end + 1, isize);
 290         u64 inline_len = actual_end - start;
 291         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 292         u64 data_len = inline_len;
 293         int ret;
 294         struct btrfs_path *path;
 295         int extent_inserted = 0;
 296         u32 extent_item_size;
 297
 298         if (compressed_size)
 299                 data_len = compressed_size;
 300
 301         if (start > 0 ||
 302             actual_end > fs_info->sectorsize ||
 303             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 304             (!compressed_size &&
 305             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 306             end + 1 < isize ||
 307             data_len > fs_info->max_inline) {
 308                 return 1;
 309         }
 310
 311         path = btrfs_alloc_path();
 312         if (!path)
 313                 return -ENOMEM;
 314
 315         trans = btrfs_join_transaction(root);
 316         if (IS_ERR(trans)) {
 317                 btrfs_free_path(path);
 318                 return PTR_ERR(trans);
 319         }
 320         trans->block_rsv = &fs_info->delalloc_block_rsv;
 321
 322         if (compressed_size && compressed_pages)
 323                 extent_item_size = btrfs_file_extent_calc_inline_size(
 324                    compressed_size);
 325         else
 326                 extent_item_size = btrfs_file_extent_calc_inline_size(
 327                     inline_len);
 328
 329         ret = __btrfs_drop_extents(trans, root, inode, path,
 330                                    start, aligned_end, NULL,
 331                                    1, 1, extent_item_size, &extent_inserted);
 332         if (ret) {
 333                 btrfs_abort_transaction(trans, ret);
 334                 goto out;
 335         }
 336
 337         if (isize > actual_end)
 338                 inline_len = min_t(u64, isize, actual_end);
 339         ret = insert_inline_extent(trans, path, extent_inserted,
 340                                    root, inode, start,
 341                                    inline_len, compressed_size,
 342                                    compress_type, compressed_pages);
 343         if (ret && ret != -ENOSPC) {
 344                 btrfs_abort_transaction(trans, ret);
 345                 goto out;
 346         } else if (ret == -ENOSPC) {
 347                 ret = 1;
 348                 goto out;
 349         }
 350
 351         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 352         btrfs_delalloc_release_metadata(BTRFS_I(inode), end + 1 - start);
 353         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 354 out:
 355         /*
 356          * Don't forget to free the reserved space, as for inlined extent
 357          * it won't count as data extent, free them directly here.
 358          * And at reserve time, it's always aligned to page size, so
 359          * just free one page here.
 360          */
 361         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 362         btrfs_free_path(path);
 363         btrfs_end_transaction(trans);
 364         return ret;
 365 }
 366
 367 struct async_extent {
 368         u64 start;
 369         u64 ram_size;
 370         u64 compressed_size;
 371         struct page **pages;
 372         unsigned long nr_pages;
 373         int compress_type;
 374         struct list_head list;
 375 };
 376
 377 struct async_cow {
 378         struct inode *inode;
 379         struct btrfs_root *root;
 380         struct page *locked_page;
 381         u64 start;
 382         u64 end;
 383         struct list_head extents;
 384         struct btrfs_work work;
 385 };
 386
 387 static noinline int add_async_extent(struct async_cow *cow,
 388                                      u64 start, u64 ram_size,
 389                                      u64 compressed_size,
 390                                      struct page **pages,
 391                                      unsigned long nr_pages,
 392                                      int compress_type)
 393 {
 394         struct async_extent *async_extent;
 395
 396         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 397         BUG_ON(!async_extent); /* -ENOMEM */
 398         async_extent->start = start;
 399         async_extent->ram_size = ram_size;
 400         async_extent->compressed_size = compressed_size;
 401         async_extent->pages = pages;
 402         async_extent->nr_pages = nr_pages;
 403         async_extent->compress_type = compress_type;
 404         list_add_tail(&async_extent->list, &cow->extents);
 405         return 0;
 406 }
 407
 408 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 409 {
 410         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 411
 412         /* force compress */
 413         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 414                 return 1;
 415         /* defrag ioctl */
 416         if (BTRFS_I(inode)->defrag_compress)
 417                 return 1;
 418         /* bad compression ratios */
 419         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 420                 return 0;
 421         if (btrfs_test_opt(fs_info, COMPRESS) ||
 422             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 423             BTRFS_I(inode)->prop_compress)
 424                 return btrfs_compress_heuristic(inode, start, end);
 425         return 0;
 426 }
 427
 428 static inline void inode_should_defrag(struct btrfs_inode *inode,
 429                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 430 {
 431         /* If this is a small write inside eof, kick off a defrag */
 432         if (num_bytes < small_write &&
 433             (start > 0 || end + 1 < inode->disk_i_size))
 434                 btrfs_add_inode_defrag(NULL, inode);
 435 }
 436
 437 /*
 438  * we create compressed extents in two phases.  The first
 439  * phase compresses a range of pages that have already been
 440  * locked (both pages and state bits are locked).
 441  *
 442  * This is done inside an ordered work queue, and the compression
 443  * is spread across many cpus.  The actual IO submission is step
 444  * two, and the ordered work queue takes care of making sure that
 445  * happens in the same order things were put onto the queue by
 446  * writepages and friends.
 447  *
 448  * If this code finds it can't get good compression, it puts an
 449  * entry onto the work queue to write the uncompressed bytes.  This
 450  * makes sure that both compressed inodes and uncompressed inodes
 451  * are written in the same order that the flusher thread sent them
 452  * down.
 453  */
 454 static noinline void compress_file_range(struct inode *inode,
 455                                         struct page *locked_page,
 456                                         u64 start, u64 end,
 457                                         struct async_cow *async_cow,
 458                                         int *num_added)
 459 {
 460         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 461         struct btrfs_root *root = BTRFS_I(inode)->root;
 462         u64 num_bytes;
 463         u64 blocksize = fs_info->sectorsize;
 464         u64 actual_end;
 465         u64 isize = i_size_read(inode);
 466         int ret = 0;
 467         struct page **pages = NULL;
 468         unsigned long nr_pages;
 469         unsigned long total_compressed = 0;
 470         unsigned long total_in = 0;
 471         int i;
 472         int will_compress;
 473         int compress_type = fs_info->compress_type;
 474         int redirty = 0;
 475
 476         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 477                         SZ_16K);
 478
 479         actual_end = min_t(u64, isize, end + 1);
 480 again:
 481         will_compress = 0;
 482         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 483         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 484         nr_pages = min_t(unsigned long, nr_pages,
 485                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 486
 487         /*
 488          * we don't want to send crud past the end of i_size through
 489          * compression, that's just a waste of CPU time.  So, if the
 490          * end of the file is before the start of our current
 491          * requested range of bytes, we bail out to the uncompressed
 492          * cleanup code that can deal with all of this.
 493          *
 494          * It isn't really the fastest way to fix things, but this is a
 495          * very uncommon corner.
 496          */
 497         if (actual_end <= start)
 498                 goto cleanup_and_bail_uncompressed;
 499
 500         total_compressed = actual_end - start;
 501
 502         /*
 503          * skip compression for a small file range(<=blocksize) that
 504          * isn't an inline extent, since it doesn't save disk space at all.
 505          */
 506         if (total_compressed <= blocksize &&
 507            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 508                 goto cleanup_and_bail_uncompressed;
 509
 510         total_compressed = min_t(unsigned long, total_compressed,
 511                         BTRFS_MAX_UNCOMPRESSED);
 512         num_bytes = ALIGN(end - start + 1, blocksize);
 513         num_bytes = max(blocksize,  num_bytes);
 514         total_in = 0;
 515         ret = 0;
 516
 517         /*
 518          * we do compression for mount -o compress and when the
 519          * inode has not been flagged as nocompress.  This flag can
 520          * change at any time if we discover bad compression ratios.
 521          */
 522         if (inode_need_compress(inode, start, end)) {
 523                 WARN_ON(pages);
 524                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 525                 if (!pages) {
 526                         /* just bail out to the uncompressed code */
 527                         nr_pages = 0;
 528                         goto cont;
 529                 }
 530
 531                 if (BTRFS_I(inode)->defrag_compress)
 532                         compress_type = BTRFS_I(inode)->defrag_compress;
 533                 else if (BTRFS_I(inode)->prop_compress)
 534                         compress_type = BTRFS_I(inode)->prop_compress;
 535
 536                 /*
 537                  * we need to call clear_page_dirty_for_io on each
 538                  * page in the range.  Otherwise applications with the file
 539                  * mmap'd can wander in and change the page contents while
 540                  * we are compressing them.
 541                  *
 542                  * If the compression fails for any reason, we set the pages
 543                  * dirty again later on.
 544                  */
 545                 extent_range_clear_dirty_for_io(inode, start, end);
 546                 redirty = 1;
 547                 ret = btrfs_compress_pages(compress_type,
 548                                            inode->i_mapping, start,
 549                                            pages,
 550                                            &nr_pages,
 551                                            &total_in,
 552                                            &total_compressed);
 553
 554                 if (!ret) {
 555                         unsigned long offset = total_compressed &
 556                                 (PAGE_SIZE - 1);
 557                         struct page *page = pages[nr_pages - 1];
 558                         char *kaddr;
 559
 560                         /* zero the tail end of the last page, we might be
 561                          * sending it down to disk
 562                          */
 563                         if (offset) {
 564                                 kaddr = kmap_atomic(page);
 565                                 memset(kaddr + offset, 0,
 566                                        PAGE_SIZE - offset);
 567                                 kunmap_atomic(kaddr);
 568                         }
 569                         will_compress = 1;
 570                 }
 571         }
 572 cont:
 573         if (start == 0) {
 574                 /* lets try to make an inline extent */
 575                 if (ret || total_in < (actual_end - start)) {
 576                         /* we didn't compress the entire range, try
 577                          * to make an uncompressed inline extent.
 578                          */
 579                         ret = cow_file_range_inline(root, inode, start, end,
 580                                             0, BTRFS_COMPRESS_NONE, NULL);
 581                 } else {
 582                         /* try making a compressed inline extent */
 583                         ret = cow_file_range_inline(root, inode, start, end,
 584                                                     total_compressed,
 585                                                     compress_type, pages);
 586                 }
 587                 if (ret <= 0) {
 588                         unsigned long clear_flags = EXTENT_DELALLOC |
 589                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG;
 590                         unsigned long page_error_op;
 591
 592                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 593                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 594
 595                         /*
 596                          * inline extent creation worked or returned error,
 597                          * we don't need to create any more async work items.
 598                          * Unlock and free up our temp pages.
 599                          */
 600                         extent_clear_unlock_delalloc(inode, start, end, end,
 601                                                      NULL, clear_flags,
 602                                                      PAGE_UNLOCK |
 603                                                      PAGE_CLEAR_DIRTY |
 604                                                      PAGE_SET_WRITEBACK |
 605                                                      page_error_op |
 606                                                      PAGE_END_WRITEBACK);
 607                         if (ret == 0)
 608                                 btrfs_free_reserved_data_space_noquota(inode,
 609                                                                start,
 610                                                                end - start + 1);
 611                         goto free_pages_out;
 612                 }
 613         }
 614
 615         if (will_compress) {
 616                 /*
 617                  * we aren't doing an inline extent round the compressed size
 618                  * up to a block size boundary so the allocator does sane
 619                  * things
 620                  */
 621                 total_compressed = ALIGN(total_compressed, blocksize);
 622
 623                 /*
 624                  * one last check to make sure the compression is really a
 625                  * win, compare the page count read with the blocks on disk,
 626                  * compression must free at least one sector size
 627                  */
 628                 total_in = ALIGN(total_in, PAGE_SIZE);
 629                 if (total_compressed + blocksize <= total_in) {
 630                         num_bytes = total_in;
 631                         *num_added += 1;
 632
 633                         /*
 634                          * The async work queues will take care of doing actual
 635                          * allocation on disk for these compressed pages, and
 636                          * will submit them to the elevator.
 637                          */
 638                         add_async_extent(async_cow, start, num_bytes,
 639                                         total_compressed, pages, nr_pages,
 640                                         compress_type);
 641
 642                         if (start + num_bytes < end) {
 643                                 start += num_bytes;
 644                                 pages = NULL;
 645                                 cond_resched();
 646                                 goto again;
 647                         }
 648                         return;
 649                 }
 650         }
 651         if (pages) {
 652                 /*
 653                  * the compression code ran but failed to make things smaller,
 654                  * free any pages it allocated and our page pointer array
 655                  */
 656                 for (i = 0; i < nr_pages; i++) {
 657                         WARN_ON(pages[i]->mapping);
 658                         put_page(pages[i]);
 659                 }
 660                 kfree(pages);
 661                 pages = NULL;
 662                 total_compressed = 0;
 663                 nr_pages = 0;
 664
 665                 /* flag the file so we don't compress in the future */
 666                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 667                     !(BTRFS_I(inode)->prop_compress)) {
 668                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 669                 }
 670         }
 671 cleanup_and_bail_uncompressed:
 672         /*
 673          * No compression, but we still need to write the pages in the file
 674          * we've been given so far.  redirty the locked page if it corresponds
 675          * to our extent and set things up for the async work queue to run
 676          * cow_file_range to do the normal delalloc dance.
 677          */
 678         if (page_offset(locked_page) >= start &&
 679             page_offset(locked_page) <= end)
 680                 __set_page_dirty_nobuffers(locked_page);
 681                 /* unlocked later on in the async handlers */
 682
 683         if (redirty)
 684                 extent_range_redirty_for_io(inode, start, end);
 685         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 686                          BTRFS_COMPRESS_NONE);
 687         *num_added += 1;
 688
 689         return;
 690
 691 free_pages_out:
 692         for (i = 0; i < nr_pages; i++) {
 693                 WARN_ON(pages[i]->mapping);
 694                 put_page(pages[i]);
 695         }
 696         kfree(pages);
 697 }
 698
 699 static void free_async_extent_pages(struct async_extent *async_extent)
 700 {
 701         int i;
 702
 703         if (!async_extent->pages)
 704                 return;
 705
 706         for (i = 0; i < async_extent->nr_pages; i++) {
 707                 WARN_ON(async_extent->pages[i]->mapping);
 708                 put_page(async_extent->pages[i]);
 709         }
 710         kfree(async_extent->pages);
 711         async_extent->nr_pages = 0;
 712         async_extent->pages = NULL;
 713 }
 714
 715 /*
 716  * phase two of compressed writeback.  This is the ordered portion
 717  * of the code, which only gets called in the order the work was
 718  * queued.  We walk all the async extents created by compress_file_range
 719  * and send them down to the disk.
 720  */
 721 static noinline void submit_compressed_extents(struct inode *inode,
 722                                               struct async_cow *async_cow)
 723 {
 724         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 725         struct async_extent *async_extent;
 726         u64 alloc_hint = 0;
 727         struct btrfs_key ins;
 728         struct extent_map *em;
 729         struct btrfs_root *root = BTRFS_I(inode)->root;
 730         struct extent_io_tree *io_tree;
 731         int ret = 0;
 732
 733 again:
 734         while (!list_empty(&async_cow->extents)) {
 735                 async_extent = list_entry(async_cow->extents.next,
 736                                           struct async_extent, list);
 737                 list_del(&async_extent->list);
 738
 739                 io_tree = &BTRFS_I(inode)->io_tree;
 740
 741 retry:
 742                 /* did the compression code fall back to uncompressed IO? */
 743                 if (!async_extent->pages) {
 744                         int page_started = 0;
 745                         unsigned long nr_written = 0;
 746
 747                         lock_extent(io_tree, async_extent->start,
 748                                          async_extent->start +
 749                                          async_extent->ram_size - 1);
 750
 751                         /* allocate blocks */
 752                         ret = cow_file_range(inode, async_cow->locked_page,
 753                                              async_extent->start,
 754                                              async_extent->start +
 755                                              async_extent->ram_size - 1,
 756                                              async_extent->start +
 757                                              async_extent->ram_size - 1,
 758                                              &page_started, &nr_written, 0,
 759                                              NULL);
 760
 761                         /* JDM XXX */
 762
 763                         /*
 764                          * if page_started, cow_file_range inserted an
 765                          * inline extent and took care of all the unlocking
 766                          * and IO for us.  Otherwise, we need to submit
 767                          * all those pages down to the drive.
 768                          */
 769                         if (!page_started && !ret)
 770                                 extent_write_locked_range(io_tree,
 771                                                   inode, async_extent->start,
 772                                                   async_extent->start +
 773                                                   async_extent->ram_size - 1,
 774                                                   btrfs_get_extent,
 775                                                   WB_SYNC_ALL);
 776                         else if (ret)
 777                                 unlock_page(async_cow->locked_page);
 778                         kfree(async_extent);
 779                         cond_resched();
 780                         continue;
 781                 }
 782
 783                 lock_extent(io_tree, async_extent->start,
 784                             async_extent->start + async_extent->ram_size - 1);
 785
 786                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 787                                            async_extent->compressed_size,
 788                                            async_extent->compressed_size,
 789                                            0, alloc_hint, &ins, 1, 1);
 790                 if (ret) {
 791                         free_async_extent_pages(async_extent);
 792
 793                         if (ret == -ENOSPC) {
 794                                 unlock_extent(io_tree, async_extent->start,
 795                                               async_extent->start +
 796                                               async_extent->ram_size - 1);
 797
 798                                 /*
 799                                  * we need to redirty the pages if we decide to
 800                                  * fallback to uncompressed IO, otherwise we
 801                                  * will not submit these pages down to lower
 802                                  * layers.
 803                                  */
 804                                 extent_range_redirty_for_io(inode,
 805                                                 async_extent->start,
 806                                                 async_extent->start +
 807                                                 async_extent->ram_size - 1);
 808
 809                                 goto retry;
 810                         }
 811                         goto out_free;
 812                 }
 813                 /*
 814                  * here we're doing allocation and writeback of the
 815                  * compressed pages
 816                  */
 817                 em = create_io_em(inode, async_extent->start,
 818                                   async_extent->ram_size, /* len */
 819                                   async_extent->start, /* orig_start */
 820                                   ins.objectid, /* block_start */
 821                                   ins.offset, /* block_len */
 822                                   ins.offset, /* orig_block_len */
 823                                   async_extent->ram_size, /* ram_bytes */
 824                                   async_extent->compress_type,
 825                                   BTRFS_ORDERED_COMPRESSED);
 826                 if (IS_ERR(em))
 827                         /* ret value is not necessary due to void function */
 828                         goto out_free_reserve;
 829                 free_extent_map(em);
 830
 831                 ret = btrfs_add_ordered_extent_compress(inode,
 832                                                 async_extent->start,
 833                                                 ins.objectid,
 834                                                 async_extent->ram_size,
 835                                                 ins.offset,
 836                                                 BTRFS_ORDERED_COMPRESSED,
 837                                                 async_extent->compress_type);
 838                 if (ret) {
 839                         btrfs_drop_extent_cache(BTRFS_I(inode),
 840                                                 async_extent->start,
 841                                                 async_extent->start +
 842                                                 async_extent->ram_size - 1, 0);
 843                         goto out_free_reserve;
 844                 }
 845                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 846
 847                 /*
 848                  * clear dirty, set writeback and unlock the pages.
 849                  */
 850                 extent_clear_unlock_delalloc(inode, async_extent->start,
 851                                 async_extent->start +
 852                                 async_extent->ram_size - 1,
 853                                 async_extent->start +
 854                                 async_extent->ram_size - 1,
 855                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 856                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 857                                 PAGE_SET_WRITEBACK);
 858                 if (btrfs_submit_compressed_write(inode,
 859                                     async_extent->start,
 860                                     async_extent->ram_size,
 861                                     ins.objectid,
 862                                     ins.offset, async_extent->pages,
 863                                     async_extent->nr_pages)) {
 864                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 865                         struct page *p = async_extent->pages[0];
 866                         const u64 start = async_extent->start;
 867                         const u64 end = start + async_extent->ram_size - 1;
 868
 869                         p->mapping = inode->i_mapping;
 870                         tree->ops->writepage_end_io_hook(p, start, end,
 871                                                          NULL, 0);
 872                         p->mapping = NULL;
 873                         extent_clear_unlock_delalloc(inode, start, end, end,
 874                                                      NULL, 0,
 875                                                      PAGE_END_WRITEBACK |
 876                                                      PAGE_SET_ERROR);
 877                         free_async_extent_pages(async_extent);
 878                 }
 879                 alloc_hint = ins.objectid + ins.offset;
 880                 kfree(async_extent);
 881                 cond_resched();
 882         }
 883         return;
 884 out_free_reserve:
 885         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 886         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 887 out_free:
 888         extent_clear_unlock_delalloc(inode, async_extent->start,
 889                                      async_extent->start +
 890                                      async_extent->ram_size - 1,
 891                                      async_extent->start +
 892                                      async_extent->ram_size - 1,
 893                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 894                                      EXTENT_DELALLOC_NEW |
 895                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 896                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 897                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 898                                      PAGE_SET_ERROR);
 899         free_async_extent_pages(async_extent);
 900         kfree(async_extent);
 901         goto again;
 902 }
 903
 904 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 905                                       u64 num_bytes)
 906 {
 907         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 908         struct extent_map *em;
 909         u64 alloc_hint = 0;
 910
 911         read_lock(&em_tree->lock);
 912         em = search_extent_mapping(em_tree, start, num_bytes);
 913         if (em) {
 914                 /*
 915                  * if block start isn't an actual block number then find the
 916                  * first block in this inode and use that as a hint.  If that
 917                  * block is also bogus then just don't worry about it.
 918                  */
 919                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 920                         free_extent_map(em);
 921                         em = search_extent_mapping(em_tree, 0, 0);
 922                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 923                                 alloc_hint = em->block_start;
 924                         if (em)
 925                                 free_extent_map(em);
 926                 } else {
 927                         alloc_hint = em->block_start;
 928                         free_extent_map(em);
 929                 }
 930         }
 931         read_unlock(&em_tree->lock);
 932
 933         return alloc_hint;
 934 }
 935
 936 /*
 937  * when extent_io.c finds a delayed allocation range in the file,
 938  * the call backs end up in this code.  The basic idea is to
 939  * allocate extents on disk for the range, and create ordered data structs
 940  * in ram to track those extents.
 941  *
 942  * locked_page is the page that writepage had locked already.  We use
 943  * it to make sure we don't do extra locks or unlocks.
 944  *
 945  * *page_started is set to one if we unlock locked_page and do everything
 946  * required to start IO on it.  It may be clean and already done with
 947  * IO when we return.
 948  */
 949 static noinline int cow_file_range(struct inode *inode,
 950                                    struct page *locked_page,
 951                                    u64 start, u64 end, u64 delalloc_end,
 952                                    int *page_started, unsigned long *nr_written,
 953                                    int unlock, struct btrfs_dedupe_hash *hash)
 954 {
 955         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 956         struct btrfs_root *root = BTRFS_I(inode)->root;
 957         u64 alloc_hint = 0;
 958         u64 num_bytes;
 959         unsigned long ram_size;
 960         u64 disk_num_bytes;
 961         u64 cur_alloc_size = 0;
 962         u64 blocksize = fs_info->sectorsize;
 963         struct btrfs_key ins;
 964         struct extent_map *em;
 965         unsigned clear_bits;
 966         unsigned long page_ops;
 967         bool extent_reserved = false;
 968         int ret = 0;
 969
 970         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 971                 WARN_ON_ONCE(1);
 972                 ret = -EINVAL;
 973                 goto out_unlock;
 974         }
 975
 976         num_bytes = ALIGN(end - start + 1, blocksize);
 977         num_bytes = max(blocksize,  num_bytes);
 978         disk_num_bytes = num_bytes;
 979
 980         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 981
 982         if (start == 0) {
 983                 /* lets try to make an inline extent */
 984                 ret = cow_file_range_inline(root, inode, start, end, 0,
 985                                         BTRFS_COMPRESS_NONE, NULL);
 986                 if (ret == 0) {
 987                         extent_clear_unlock_delalloc(inode, start, end,
 988                                      delalloc_end, NULL,
 989                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 990                                      EXTENT_DELALLOC_NEW |
 991                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 992                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 993                                      PAGE_END_WRITEBACK);
 994                         btrfs_free_reserved_data_space_noquota(inode, start,
 995                                                 end - start + 1);
 996                         *nr_written = *nr_written +
 997                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 998                         *page_started = 1;
 999                         goto out;
1000                 } else if (ret < 0) {
1001                         goto out_unlock;
1002                 }
1003         }
1004
1005         BUG_ON(disk_num_bytes >
1006                btrfs_super_total_bytes(fs_info->super_copy));
1007
1008         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1009         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1010                         start + num_bytes - 1, 0);
1011
1012         while (disk_num_bytes > 0) {
1013                 cur_alloc_size = disk_num_bytes;
1014                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1015                                            fs_info->sectorsize, 0, alloc_hint,
1016                                            &ins, 1, 1);
1017                 if (ret < 0)
1018                         goto out_unlock;
1019                 cur_alloc_size = ins.offset;
1020                 extent_reserved = true;
1021
1022                 ram_size = ins.offset;
1023                 em = create_io_em(inode, start, ins.offset, /* len */
1024                                   start, /* orig_start */
1025                                   ins.objectid, /* block_start */
1026                                   ins.offset, /* block_len */
1027                                   ins.offset, /* orig_block_len */
1028                                   ram_size, /* ram_bytes */
1029                                   BTRFS_COMPRESS_NONE, /* compress_type */
1030                                   BTRFS_ORDERED_REGULAR /* type */);
1031                 if (IS_ERR(em)) {
1032                         ret = PTR_ERR(em);
1033                         goto out_reserve;
1034                 }
1035                 free_extent_map(em);
1036
1037                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1038                                                ram_size, cur_alloc_size, 0);
1039                 if (ret)
1040                         goto out_drop_extent_cache;
1041
1042                 if (root->root_key.objectid ==
1043                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1044                         ret = btrfs_reloc_clone_csums(inode, start,
1045                                                       cur_alloc_size);
1046                         /*
1047                          * Only drop cache here, and process as normal.
1048                          *
1049                          * We must not allow extent_clear_unlock_delalloc()
1050                          * at out_unlock label to free meta of this ordered
1051                          * extent, as its meta should be freed by
1052                          * btrfs_finish_ordered_io().
1053                          *
1054                          * So we must continue until @start is increased to
1055                          * skip current ordered extent.
1056                          */
1057                         if (ret)
1058                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1059                                                 start + ram_size - 1, 0);
1060                 }
1061
1062                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1063
1064                 /* we're not doing compressed IO, don't unlock the first
1065                  * page (which the caller expects to stay locked), don't
1066                  * clear any dirty bits and don't set any writeback bits
1067                  *
1068                  * Do set the Private2 bit so we know this page was properly
1069                  * setup for writepage
1070                  */
1071                 page_ops = unlock ? PAGE_UNLOCK : 0;
1072                 page_ops |= PAGE_SET_PRIVATE2;
1073
1074                 extent_clear_unlock_delalloc(inode, start,
1075                                              start + ram_size - 1,
1076                                              delalloc_end, locked_page,
1077                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1078                                              page_ops);
1079                 if (disk_num_bytes < cur_alloc_size)
1080                         disk_num_bytes = 0;
1081                 else
1082                         disk_num_bytes -= cur_alloc_size;
1083                 num_bytes -= cur_alloc_size;
1084                 alloc_hint = ins.objectid + ins.offset;
1085                 start += cur_alloc_size;
1086                 extent_reserved = false;
1087
1088                 /*
1089                  * btrfs_reloc_clone_csums() error, since start is increased
1090                  * extent_clear_unlock_delalloc() at out_unlock label won't
1091                  * free metadata of current ordered extent, we're OK to exit.
1092                  */
1093                 if (ret)
1094                         goto out_unlock;
1095         }
1096 out:
1097         return ret;
1098
1099 out_drop_extent_cache:
1100         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1101 out_reserve:
1102         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1103         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1104 out_unlock:
1105         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1106                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1107         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1108                 PAGE_END_WRITEBACK;
1109         /*
1110          * If we reserved an extent for our delalloc range (or a subrange) and
1111          * failed to create the respective ordered extent, then it means that
1112          * when we reserved the extent we decremented the extent's size from
1113          * the data space_info's bytes_may_use counter and incremented the
1114          * space_info's bytes_reserved counter by the same amount. We must make
1115          * sure extent_clear_unlock_delalloc() does not try to decrement again
1116          * the data space_info's bytes_may_use counter, therefore we do not pass
1117          * it the flag EXTENT_CLEAR_DATA_RESV.
1118          */
1119         if (extent_reserved) {
1120                 extent_clear_unlock_delalloc(inode, start,
1121                                              start + cur_alloc_size,
1122                                              start + cur_alloc_size,
1123                                              locked_page,
1124                                              clear_bits,
1125                                              page_ops);
1126                 start += cur_alloc_size;
1127                 if (start >= end)
1128                         goto out;
1129         }
1130         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1131                                      locked_page,
1132                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1133                                      page_ops);
1134         goto out;
1135 }
1136
1137 /*
1138  * work queue call back to started compression on a file and pages
1139  */
1140 static noinline void async_cow_start(struct btrfs_work *work)
1141 {
1142         struct async_cow *async_cow;
1143         int num_added = 0;
1144         async_cow = container_of(work, struct async_cow, work);
1145
1146         compress_file_range(async_cow->inode, async_cow->locked_page,
1147                             async_cow->start, async_cow->end, async_cow,
1148                             &num_added);
1149         if (num_added == 0) {
1150                 btrfs_add_delayed_iput(async_cow->inode);
1151                 async_cow->inode = NULL;
1152         }
1153 }
1154
1155 /*
1156  * work queue call back to submit previously compressed pages
1157  */
1158 static noinline void async_cow_submit(struct btrfs_work *work)
1159 {
1160         struct btrfs_fs_info *fs_info;
1161         struct async_cow *async_cow;
1162         struct btrfs_root *root;
1163         unsigned long nr_pages;
1164
1165         async_cow = container_of(work, struct async_cow, work);
1166
1167         root = async_cow->root;
1168         fs_info = root->fs_info;
1169         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1170                 PAGE_SHIFT;
1171
1172         /*
1173          * atomic_sub_return implies a barrier for waitqueue_active
1174          */
1175         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1176             5 * SZ_1M &&
1177             waitqueue_active(&fs_info->async_submit_wait))
1178                 wake_up(&fs_info->async_submit_wait);
1179
1180         if (async_cow->inode)
1181                 submit_compressed_extents(async_cow->inode, async_cow);
1182 }
1183
1184 static noinline void async_cow_free(struct btrfs_work *work)
1185 {
1186         struct async_cow *async_cow;
1187         async_cow = container_of(work, struct async_cow, work);
1188         if (async_cow->inode)
1189                 btrfs_add_delayed_iput(async_cow->inode);
1190         kfree(async_cow);
1191 }
1192
1193 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1194                                 u64 start, u64 end, int *page_started,
1195                                 unsigned long *nr_written)
1196 {
1197         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1198         struct async_cow *async_cow;
1199         struct btrfs_root *root = BTRFS_I(inode)->root;
1200         unsigned long nr_pages;
1201         u64 cur_end;
1202
1203         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1204                          1, 0, NULL, GFP_NOFS);
1205         while (start < end) {
1206                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1207                 BUG_ON(!async_cow); /* -ENOMEM */
1208                 async_cow->inode = igrab(inode);
1209                 async_cow->root = root;
1210                 async_cow->locked_page = locked_page;
1211                 async_cow->start = start;
1212
1213                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1214                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1215                         cur_end = end;
1216                 else
1217                         cur_end = min(end, start + SZ_512K - 1);
1218
1219                 async_cow->end = cur_end;
1220                 INIT_LIST_HEAD(&async_cow->extents);
1221
1222                 btrfs_init_work(&async_cow->work,
1223                                 btrfs_delalloc_helper,
1224                                 async_cow_start, async_cow_submit,
1225                                 async_cow_free);
1226
1227                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1228                         PAGE_SHIFT;
1229                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1230
1231                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1232
1233                 while (atomic_read(&fs_info->async_submit_draining) &&
1234                        atomic_read(&fs_info->async_delalloc_pages)) {
1235                         wait_event(fs_info->async_submit_wait,
1236                                    (atomic_read(&fs_info->async_delalloc_pages) ==
1237                                     0));
1238                 }
1239
1240                 *nr_written += nr_pages;
1241                 start = cur_end + 1;
1242         }
1243         *page_started = 1;
1244         return 0;
1245 }
1246
1247 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1248                                         u64 bytenr, u64 num_bytes)
1249 {
1250         int ret;
1251         struct btrfs_ordered_sum *sums;
1252         LIST_HEAD(list);
1253
1254         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1255                                        bytenr + num_bytes - 1, &list, 0);
1256         if (ret == 0 && list_empty(&list))
1257                 return 0;
1258
1259         while (!list_empty(&list)) {
1260                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1261                 list_del(&sums->list);
1262                 kfree(sums);
1263         }
1264         if (ret < 0)
1265                 return ret;
1266         return 1;
1267 }
1268
1269 /*
1270  * when nowcow writeback call back.  This checks for snapshots or COW copies
1271  * of the extents that exist in the file, and COWs the file as required.
1272  *
1273  * If no cow copies or snapshots exist, we write directly to the existing
1274  * blocks on disk
1275  */
1276 static noinline int run_delalloc_nocow(struct inode *inode,
1277                                        struct page *locked_page,
1278                               u64 start, u64 end, int *page_started, int force,
1279                               unsigned long *nr_written)
1280 {
1281         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1282         struct btrfs_root *root = BTRFS_I(inode)->root;
1283         struct extent_buffer *leaf;
1284         struct btrfs_path *path;
1285         struct btrfs_file_extent_item *fi;
1286         struct btrfs_key found_key;
1287         struct extent_map *em;
1288         u64 cow_start;
1289         u64 cur_offset;
1290         u64 extent_end;
1291         u64 extent_offset;
1292         u64 disk_bytenr;
1293         u64 num_bytes;
1294         u64 disk_num_bytes;
1295         u64 ram_bytes;
1296         int extent_type;
1297         int ret, err;
1298         int type;
1299         int nocow;
1300         int check_prev = 1;
1301         bool nolock;
1302         u64 ino = btrfs_ino(BTRFS_I(inode));
1303
1304         path = btrfs_alloc_path();
1305         if (!path) {
1306                 extent_clear_unlock_delalloc(inode, start, end, end,
1307                                              locked_page,
1308                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1309                                              EXTENT_DO_ACCOUNTING |
1310                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1311                                              PAGE_CLEAR_DIRTY |
1312                                              PAGE_SET_WRITEBACK |
1313                                              PAGE_END_WRITEBACK);
1314                 return -ENOMEM;
1315         }
1316
1317         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1318
1319         cow_start = (u64)-1;
1320         cur_offset = start;
1321         while (1) {
1322                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1323                                                cur_offset, 0);
1324                 if (ret < 0)
1325                         goto error;
1326                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1327                         leaf = path->nodes[0];
1328                         btrfs_item_key_to_cpu(leaf, &found_key,
1329                                               path->slots[0] - 1);
1330                         if (found_key.objectid == ino &&
1331                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1332                                 path->slots[0]--;
1333                 }
1334                 check_prev = 0;
1335 next_slot:
1336                 leaf = path->nodes[0];
1337                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1338                         ret = btrfs_next_leaf(root, path);
1339                         if (ret < 0) {
1340                                 if (cow_start != (u64)-1)
1341                                         cur_offset = cow_start;
1342                                 goto error;
1343                         }
1344                         if (ret > 0)
1345                                 break;
1346                         leaf = path->nodes[0];
1347                 }
1348
1349                 nocow = 0;
1350                 disk_bytenr = 0;
1351                 num_bytes = 0;
1352                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1353
1354                 if (found_key.objectid > ino)
1355                         break;
1356                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1357                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1358                         path->slots[0]++;
1359                         goto next_slot;
1360                 }
1361                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1362                     found_key.offset > end)
1363                         break;
1364
1365                 if (found_key.offset > cur_offset) {
1366                         extent_end = found_key.offset;
1367                         extent_type = 0;
1368                         goto out_check;
1369                 }
1370
1371                 fi = btrfs_item_ptr(leaf, path->slots[0],
1372                                     struct btrfs_file_extent_item);
1373                 extent_type = btrfs_file_extent_type(leaf, fi);
1374
1375                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1376                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1377                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1378                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1379                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1380                         extent_end = found_key.offset +
1381                                 btrfs_file_extent_num_bytes(leaf, fi);
1382                         disk_num_bytes =
1383                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1384                         if (extent_end <= start) {
1385                                 path->slots[0]++;
1386                                 goto next_slot;
1387                         }
1388                         if (disk_bytenr == 0)
1389                                 goto out_check;
1390                         if (btrfs_file_extent_compression(leaf, fi) ||
1391                             btrfs_file_extent_encryption(leaf, fi) ||
1392                             btrfs_file_extent_other_encoding(leaf, fi))
1393                                 goto out_check;
1394                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1395                                 goto out_check;
1396                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1397                                 goto out_check;
1398                         ret = btrfs_cross_ref_exist(root, ino,
1399                                                     found_key.offset -
1400                                                     extent_offset, disk_bytenr);
1401                         if (ret) {
1402                                 /*
1403                                  * ret could be -EIO if the above fails to read
1404                                  * metadata.
1405                                  */
1406                                 if (ret < 0) {
1407                                         if (cow_start != (u64)-1)
1408                                                 cur_offset = cow_start;
1409                                         goto error;
1410                                 }
1411
1412                                 WARN_ON_ONCE(nolock);
1413                                 goto out_check;
1414                         }
1415                         disk_bytenr += extent_offset;
1416                         disk_bytenr += cur_offset - found_key.offset;
1417                         num_bytes = min(end + 1, extent_end) - cur_offset;
1418                         /*
1419                          * if there are pending snapshots for this root,
1420                          * we fall into common COW way.
1421                          */
1422                         if (!nolock) {
1423                                 err = btrfs_start_write_no_snapshotting(root);
1424                                 if (!err)
1425                                         goto out_check;
1426                         }
1427                         /*
1428                          * force cow if csum exists in the range.
1429                          * this ensure that csum for a given extent are
1430                          * either valid or do not exist.
1431                          */
1432                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1433                                                   num_bytes);
1434                         if (ret) {
1435                                 if (!nolock)
1436                                         btrfs_end_write_no_snapshotting(root);
1437
1438                                 /*
1439                                  * ret could be -EIO if the above fails to read
1440                                  * metadata.
1441                                  */
1442                                 if (ret < 0) {
1443                                         if (cow_start != (u64)-1)
1444                                                 cur_offset = cow_start;
1445                                         goto error;
1446                                 }
1447                                 WARN_ON_ONCE(nolock);
1448                                 goto out_check;
1449                         }
1450                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr)) {
1451                                 if (!nolock)
1452                                         btrfs_end_write_no_snapshotting(root);
1453                                 goto out_check;
1454                         }
1455                         nocow = 1;
1456                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1457                         extent_end = found_key.offset +
1458                                 btrfs_file_extent_inline_len(leaf,
1459                                                      path->slots[0], fi);
1460                         extent_end = ALIGN(extent_end,
1461                                            fs_info->sectorsize);
1462                 } else {
1463                         BUG_ON(1);
1464                 }
1465 out_check:
1466                 if (extent_end <= start) {
1467                         path->slots[0]++;
1468                         if (!nolock && nocow)
1469                                 btrfs_end_write_no_snapshotting(root);
1470                         if (nocow)
1471                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1472                         goto next_slot;
1473                 }
1474                 if (!nocow) {
1475                         if (cow_start == (u64)-1)
1476                                 cow_start = cur_offset;
1477                         cur_offset = extent_end;
1478                         if (cur_offset > end)
1479                                 break;
1480                         path->slots[0]++;
1481                         goto next_slot;
1482                 }
1483
1484                 btrfs_release_path(path);
1485                 if (cow_start != (u64)-1) {
1486                         ret = cow_file_range(inode, locked_page,
1487                                              cow_start, found_key.offset - 1,
1488                                              end, page_started, nr_written, 1,
1489                                              NULL);
1490                         if (ret) {
1491                                 if (!nolock && nocow)
1492                                         btrfs_end_write_no_snapshotting(root);
1493                                 if (nocow)
1494                                         btrfs_dec_nocow_writers(fs_info,
1495                                                                 disk_bytenr);
1496                                 goto error;
1497                         }
1498                         cow_start = (u64)-1;
1499                 }
1500
1501                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1502                         u64 orig_start = found_key.offset - extent_offset;
1503
1504                         em = create_io_em(inode, cur_offset, num_bytes,
1505                                           orig_start,
1506                                           disk_bytenr, /* block_start */
1507                                           num_bytes, /* block_len */
1508                                           disk_num_bytes, /* orig_block_len */
1509                                           ram_bytes, BTRFS_COMPRESS_NONE,
1510                                           BTRFS_ORDERED_PREALLOC);
1511                         if (IS_ERR(em)) {
1512                                 if (!nolock && nocow)
1513                                         btrfs_end_write_no_snapshotting(root);
1514                                 if (nocow)
1515                                         btrfs_dec_nocow_writers(fs_info,
1516                                                                 disk_bytenr);
1517                                 ret = PTR_ERR(em);
1518                                 goto error;
1519                         }
1520                         free_extent_map(em);
1521                 }
1522
1523                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1524                         type = BTRFS_ORDERED_PREALLOC;
1525                 } else {
1526                         type = BTRFS_ORDERED_NOCOW;
1527                 }
1528
1529                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1530                                                num_bytes, num_bytes, type);
1531                 if (nocow)
1532                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1533                 BUG_ON(ret); /* -ENOMEM */
1534
1535                 if (root->root_key.objectid ==
1536                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1537                         /*
1538                          * Error handled later, as we must prevent
1539                          * extent_clear_unlock_delalloc() in error handler
1540                          * from freeing metadata of created ordered extent.
1541                          */
1542                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1543                                                       num_bytes);
1544
1545                 extent_clear_unlock_delalloc(inode, cur_offset,
1546                                              cur_offset + num_bytes - 1, end,
1547                                              locked_page, EXTENT_LOCKED |
1548                                              EXTENT_DELALLOC |
1549                                              EXTENT_CLEAR_DATA_RESV,
1550                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1551
1552                 if (!nolock && nocow)
1553                         btrfs_end_write_no_snapshotting(root);
1554                 cur_offset = extent_end;
1555
1556                 /*
1557                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1558                  * handler, as metadata for created ordered extent will only
1559                  * be freed by btrfs_finish_ordered_io().
1560                  */
1561                 if (ret)
1562                         goto error;
1563                 if (cur_offset > end)
1564                         break;
1565         }
1566         btrfs_release_path(path);
1567
1568         if (cur_offset <= end && cow_start == (u64)-1)
1569                 cow_start = cur_offset;
1570
1571         if (cow_start != (u64)-1) {
1572                 cur_offset = end;
1573                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1574                                      page_started, nr_written, 1, NULL);
1575                 if (ret)
1576                         goto error;
1577         }
1578
1579 error:
1580         if (ret && cur_offset < end)
1581                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1582                                              locked_page, EXTENT_LOCKED |
1583                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1584                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1585                                              PAGE_CLEAR_DIRTY |
1586                                              PAGE_SET_WRITEBACK |
1587                                              PAGE_END_WRITEBACK);
1588         btrfs_free_path(path);
1589         return ret;
1590 }
1591
1592 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1593 {
1594
1595         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1596             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1597                 return 0;
1598
1599         /*
1600          * @defrag_bytes is a hint value, no spinlock held here,
1601          * if is not zero, it means the file is defragging.
1602          * Force cow if given extent needs to be defragged.
1603          */
1604         if (BTRFS_I(inode)->defrag_bytes &&
1605             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1606                            EXTENT_DEFRAG, 0, NULL))
1607                 return 1;
1608
1609         return 0;
1610 }
1611
1612 /*
1613  * extent_io.c call back to do delayed allocation processing
1614  */
1615 static int run_delalloc_range(void *private_data, struct page *locked_page,
1616                               u64 start, u64 end, int *page_started,
1617                               unsigned long *nr_written)
1618 {
1619         struct inode *inode = private_data;
1620         int ret;
1621         int force_cow = need_force_cow(inode, start, end);
1622
1623         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1624                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1625                                          page_started, 1, nr_written);
1626         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1627                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1628                                          page_started, 0, nr_written);
1629         } else if (!inode_need_compress(inode, start, end)) {
1630                 ret = cow_file_range(inode, locked_page, start, end, end,
1631                                       page_started, nr_written, 1, NULL);
1632         } else {
1633                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1634                         &BTRFS_I(inode)->runtime_flags);
1635                 ret = cow_file_range_async(inode, locked_page, start, end,
1636                                            page_started, nr_written);
1637         }
1638         if (ret)
1639                 btrfs_cleanup_ordered_extents(inode, start, end - start + 1);
1640         return ret;
1641 }
1642
1643 static void btrfs_split_extent_hook(void *private_data,
1644                                     struct extent_state *orig, u64 split)
1645 {
1646         struct inode *inode = private_data;
1647         u64 size;
1648
1649         /* not delalloc, ignore it */
1650         if (!(orig->state & EXTENT_DELALLOC))
1651                 return;
1652
1653         size = orig->end - orig->start + 1;
1654         if (size > BTRFS_MAX_EXTENT_SIZE) {
1655                 u32 num_extents;
1656                 u64 new_size;
1657
1658                 /*
1659                  * See the explanation in btrfs_merge_extent_hook, the same
1660                  * applies here, just in reverse.
1661                  */
1662                 new_size = orig->end - split + 1;
1663                 num_extents = count_max_extents(new_size);
1664                 new_size = split - orig->start;
1665                 num_extents += count_max_extents(new_size);
1666                 if (count_max_extents(size) >= num_extents)
1667                         return;
1668         }
1669
1670         spin_lock(&BTRFS_I(inode)->lock);
1671         BTRFS_I(inode)->outstanding_extents++;
1672         spin_unlock(&BTRFS_I(inode)->lock);
1673 }
1674
1675 /*
1676  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1677  * extents so we can keep track of new extents that are just merged onto old
1678  * extents, such as when we are doing sequential writes, so we can properly
1679  * account for the metadata space we'll need.
1680  */
1681 static void btrfs_merge_extent_hook(void *private_data,
1682                                     struct extent_state *new,
1683                                     struct extent_state *other)
1684 {
1685         struct inode *inode = private_data;
1686         u64 new_size, old_size;
1687         u32 num_extents;
1688
1689         /* not delalloc, ignore it */
1690         if (!(other->state & EXTENT_DELALLOC))
1691                 return;
1692
1693         if (new->start > other->start)
1694                 new_size = new->end - other->start + 1;
1695         else
1696                 new_size = other->end - new->start + 1;
1697
1698         /* we're not bigger than the max, unreserve the space and go */
1699         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1700                 spin_lock(&BTRFS_I(inode)->lock);
1701                 BTRFS_I(inode)->outstanding_extents--;
1702                 spin_unlock(&BTRFS_I(inode)->lock);
1703                 return;
1704         }
1705
1706         /*
1707          * We have to add up either side to figure out how many extents were
1708          * accounted for before we merged into one big extent.  If the number of
1709          * extents we accounted for is <= the amount we need for the new range
1710          * then we can return, otherwise drop.  Think of it like this
1711          *
1712          * [ 4k][MAX_SIZE]
1713          *
1714          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1715          * need 2 outstanding extents, on one side we have 1 and the other side
1716          * we have 1 so they are == and we can return.  But in this case
1717          *
1718          * [MAX_SIZE+4k][MAX_SIZE+4k]
1719          *
1720          * Each range on their own accounts for 2 extents, but merged together
1721          * they are only 3 extents worth of accounting, so we need to drop in
1722          * this case.
1723          */
1724         old_size = other->end - other->start + 1;
1725         num_extents = count_max_extents(old_size);
1726         old_size = new->end - new->start + 1;
1727         num_extents += count_max_extents(old_size);
1728         if (count_max_extents(new_size) >= num_extents)
1729                 return;
1730
1731         spin_lock(&BTRFS_I(inode)->lock);
1732         BTRFS_I(inode)->outstanding_extents--;
1733         spin_unlock(&BTRFS_I(inode)->lock);
1734 }
1735
1736 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1737                                       struct inode *inode)
1738 {
1739         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1740
1741         spin_lock(&root->delalloc_lock);
1742         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1743                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1744                               &root->delalloc_inodes);
1745                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1746                         &BTRFS_I(inode)->runtime_flags);
1747                 root->nr_delalloc_inodes++;
1748                 if (root->nr_delalloc_inodes == 1) {
1749                         spin_lock(&fs_info->delalloc_root_lock);
1750                         BUG_ON(!list_empty(&root->delalloc_root));
1751                         list_add_tail(&root->delalloc_root,
1752                                       &fs_info->delalloc_roots);
1753                         spin_unlock(&fs_info->delalloc_root_lock);
1754                 }
1755         }
1756         spin_unlock(&root->delalloc_lock);
1757 }
1758
1759
1760 void __btrfs_del_delalloc_inode(struct btrfs_root *root,
1761                                 struct btrfs_inode *inode)
1762 {
1763         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1764
1765         if (!list_empty(&inode->delalloc_inodes)) {
1766                 list_del_init(&inode->delalloc_inodes);
1767                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1768                           &inode->runtime_flags);
1769                 root->nr_delalloc_inodes--;
1770                 if (!root->nr_delalloc_inodes) {
1771                         spin_lock(&fs_info->delalloc_root_lock);
1772                         BUG_ON(list_empty(&root->delalloc_root));
1773                         list_del_init(&root->delalloc_root);
1774                         spin_unlock(&fs_info->delalloc_root_lock);
1775                 }
1776         }
1777 }
1778
1779 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1780                                      struct btrfs_inode *inode)
1781 {
1782         spin_lock(&root->delalloc_lock);
1783         __btrfs_del_delalloc_inode(root, inode);
1784         spin_unlock(&root->delalloc_lock);
1785 }
1786
1787 /*
1788  * extent_io.c set_bit_hook, used to track delayed allocation
1789  * bytes in this file, and to maintain the list of inodes that
1790  * have pending delalloc work to be done.
1791  */
1792 static void btrfs_set_bit_hook(void *private_data,
1793                                struct extent_state *state, unsigned *bits)
1794 {
1795         struct inode *inode = private_data;
1796
1797         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1798
1799         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1800                 WARN_ON(1);
1801         /*
1802          * set_bit and clear bit hooks normally require _irqsave/restore
1803          * but in this case, we are only testing for the DELALLOC
1804          * bit, which is only set or cleared with irqs on
1805          */
1806         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1807                 struct btrfs_root *root = BTRFS_I(inode)->root;
1808                 u64 len = state->end + 1 - state->start;
1809                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1810
1811                 if (*bits & EXTENT_FIRST_DELALLOC) {
1812                         *bits &= ~EXTENT_FIRST_DELALLOC;
1813                 } else {
1814                         spin_lock(&BTRFS_I(inode)->lock);
1815                         BTRFS_I(inode)->outstanding_extents++;
1816                         spin_unlock(&BTRFS_I(inode)->lock);
1817                 }
1818
1819                 /* For sanity tests */
1820                 if (btrfs_is_testing(fs_info))
1821                         return;
1822
1823                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1824                                          fs_info->delalloc_batch);
1825                 spin_lock(&BTRFS_I(inode)->lock);
1826                 BTRFS_I(inode)->delalloc_bytes += len;
1827                 if (*bits & EXTENT_DEFRAG)
1828                         BTRFS_I(inode)->defrag_bytes += len;
1829                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1830                                          &BTRFS_I(inode)->runtime_flags))
1831                         btrfs_add_delalloc_inodes(root, inode);
1832                 spin_unlock(&BTRFS_I(inode)->lock);
1833         }
1834
1835         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1836             (*bits & EXTENT_DELALLOC_NEW)) {
1837                 spin_lock(&BTRFS_I(inode)->lock);
1838                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1839                         state->start;
1840                 spin_unlock(&BTRFS_I(inode)->lock);
1841         }
1842 }
1843
1844 /*
1845  * extent_io.c clear_bit_hook, see set_bit_hook for why
1846  */
1847 static void btrfs_clear_bit_hook(void *private_data,
1848                                  struct extent_state *state,
1849                                  unsigned *bits)
1850 {
1851         struct btrfs_inode *inode = BTRFS_I((struct inode *)private_data);
1852         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1853         u64 len = state->end + 1 - state->start;
1854         u32 num_extents = count_max_extents(len);
1855
1856         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1857                 spin_lock(&inode->lock);
1858                 inode->defrag_bytes -= len;
1859                 spin_unlock(&inode->lock);
1860         }
1861
1862         /*
1863          * set_bit and clear bit hooks normally require _irqsave/restore
1864          * but in this case, we are only testing for the DELALLOC
1865          * bit, which is only set or cleared with irqs on
1866          */
1867         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1868                 struct btrfs_root *root = inode->root;
1869                 bool do_list = !btrfs_is_free_space_inode(inode);
1870
1871                 if (*bits & EXTENT_FIRST_DELALLOC) {
1872                         *bits &= ~EXTENT_FIRST_DELALLOC;
1873                 } else if (!(*bits & EXTENT_CLEAR_META_RESV)) {
1874                         spin_lock(&inode->lock);
1875                         inode->outstanding_extents -= num_extents;
1876                         spin_unlock(&inode->lock);
1877                 }
1878
1879                 /*
1880                  * We don't reserve metadata space for space cache inodes so we
1881                  * don't need to call dellalloc_release_metadata if there is an
1882                  * error.
1883                  */
1884                 if (*bits & EXTENT_CLEAR_META_RESV &&
1885                     root != fs_info->tree_root)
1886                         btrfs_delalloc_release_metadata(inode, len);
1887
1888                 /* For sanity tests. */
1889                 if (btrfs_is_testing(fs_info))
1890                         return;
1891
1892                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1893                     do_list && !(state->state & EXTENT_NORESERVE) &&
1894                     (*bits & EXTENT_CLEAR_DATA_RESV))
1895                         btrfs_free_reserved_data_space_noquota(
1896                                         &inode->vfs_inode,
1897                                         state->start, len);
1898
1899                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1900                                          fs_info->delalloc_batch);
1901                 spin_lock(&inode->lock);
1902                 inode->delalloc_bytes -= len;
1903                 if (do_list && inode->delalloc_bytes == 0 &&
1904                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1905                                         &inode->runtime_flags))
1906                         btrfs_del_delalloc_inode(root, inode);
1907                 spin_unlock(&inode->lock);
1908         }
1909
1910         if ((state->state & EXTENT_DELALLOC_NEW) &&
1911             (*bits & EXTENT_DELALLOC_NEW)) {
1912                 spin_lock(&inode->lock);
1913                 ASSERT(inode->new_delalloc_bytes >= len);
1914                 inode->new_delalloc_bytes -= len;
1915                 spin_unlock(&inode->lock);
1916         }
1917 }
1918
1919 /*
1920  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1921  * we don't create bios that span stripes or chunks
1922  *
1923  * return 1 if page cannot be merged to bio
1924  * return 0 if page can be merged to bio
1925  * return error otherwise
1926  */
1927 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1928                          size_t size, struct bio *bio,
1929                          unsigned long bio_flags)
1930 {
1931         struct inode *inode = page->mapping->host;
1932         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1933         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1934         u64 length = 0;
1935         u64 map_length;
1936         int ret;
1937
1938         if (bio_flags & EXTENT_BIO_COMPRESSED)
1939                 return 0;
1940
1941         length = bio->bi_iter.bi_size;
1942         map_length = length;
1943         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1944                               NULL, 0);
1945         if (ret < 0)
1946                 return ret;
1947         if (map_length < length + size)
1948                 return 1;
1949         return 0;
1950 }
1951
1952 /*
1953  * in order to insert checksums into the metadata in large chunks,
1954  * we wait until bio submission time.   All the pages in the bio are
1955  * checksummed and sums are attached onto the ordered extent record.
1956  *
1957  * At IO completion time the cums attached on the ordered extent record
1958  * are inserted into the btree
1959  */
1960 static blk_status_t __btrfs_submit_bio_start(void *private_data, struct bio *bio,
1961                                     int mirror_num, unsigned long bio_flags,
1962                                     u64 bio_offset)
1963 {
1964         struct inode *inode = private_data;
1965         blk_status_t ret = 0;
1966
1967         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1968         BUG_ON(ret); /* -ENOMEM */
1969         return 0;
1970 }
1971
1972 /*
1973  * in order to insert checksums into the metadata in large chunks,
1974  * we wait until bio submission time.   All the pages in the bio are
1975  * checksummed and sums are attached onto the ordered extent record.
1976  *
1977  * At IO completion time the cums attached on the ordered extent record
1978  * are inserted into the btree
1979  */
1980 static blk_status_t __btrfs_submit_bio_done(void *private_data, struct bio *bio,
1981                           int mirror_num, unsigned long bio_flags,
1982                           u64 bio_offset)
1983 {
1984         struct inode *inode = private_data;
1985         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1986         blk_status_t ret;
1987
1988         ret = btrfs_map_bio(fs_info, bio, mirror_num, 1);
1989         if (ret) {
1990                 bio->bi_status = ret;
1991                 bio_endio(bio);
1992         }
1993         return ret;
1994 }
1995
1996 /*
1997  * extent_io.c submission hook. This does the right thing for csum calculation
1998  * on write, or reading the csums from the tree before a read
1999  */
2000 static blk_status_t btrfs_submit_bio_hook(void *private_data, struct bio *bio,
2001                                  int mirror_num, unsigned long bio_flags,
2002                                  u64 bio_offset)
2003 {
2004         struct inode *inode = private_data;
2005         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2006         struct btrfs_root *root = BTRFS_I(inode)->root;
2007         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
2008         blk_status_t ret = 0;
2009         int skip_sum;
2010         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
2011
2012         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
2013
2014         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
2015                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
2016
2017         if (bio_op(bio) != REQ_OP_WRITE) {
2018                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
2019                 if (ret)
2020                         goto out;
2021
2022                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
2023                         ret = btrfs_submit_compressed_read(inode, bio,
2024                                                            mirror_num,
2025                                                            bio_flags);
2026                         goto out;
2027                 } else if (!skip_sum) {
2028                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
2029                         if (ret)
2030                                 goto out;
2031                 }
2032                 goto mapit;
2033         } else if (async && !skip_sum) {
2034                 /* csum items have already been cloned */
2035                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2036                         goto mapit;
2037                 /* we're doing a write, do the async checksumming */
2038                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2039                                           bio_offset, inode,
2040                                           __btrfs_submit_bio_start,
2041                                           __btrfs_submit_bio_done);
2042                 goto out;
2043         } else if (!skip_sum) {
2044                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2045                 if (ret)
2046                         goto out;
2047         }
2048
2049 mapit:
2050         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
2051
2052 out:
2053         if (ret) {
2054                 bio->bi_status = ret;
2055                 bio_endio(bio);
2056         }
2057         return ret;
2058 }
2059
2060 /*
2061  * given a list of ordered sums record them in the inode.  This happens
2062  * at IO completion time based on sums calculated at bio submission time.
2063  */
2064 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2065                              struct inode *inode, struct list_head *list)
2066 {
2067         struct btrfs_ordered_sum *sum;
2068
2069         list_for_each_entry(sum, list, list) {
2070                 trans->adding_csums = 1;
2071                 btrfs_csum_file_blocks(trans,
2072                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2073                 trans->adding_csums = 0;
2074         }
2075         return 0;
2076 }
2077
2078 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2079                               struct extent_state **cached_state, int dedupe)
2080 {
2081         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
2082         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2083                                    cached_state);
2084 }
2085
2086 /* see btrfs_writepage_start_hook for details on why this is required */
2087 struct btrfs_writepage_fixup {
2088         struct page *page;
2089         struct btrfs_work work;
2090 };
2091
2092 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2093 {
2094         struct btrfs_writepage_fixup *fixup;
2095         struct btrfs_ordered_extent *ordered;
2096         struct extent_state *cached_state = NULL;
2097         struct extent_changeset *data_reserved = NULL;
2098         struct page *page;
2099         struct inode *inode;
2100         u64 page_start;
2101         u64 page_end;
2102         int ret;
2103
2104         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2105         page = fixup->page;
2106 again:
2107         lock_page(page);
2108         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2109                 ClearPageChecked(page);
2110                 goto out_page;
2111         }
2112
2113         inode = page->mapping->host;
2114         page_start = page_offset(page);
2115         page_end = page_offset(page) + PAGE_SIZE - 1;
2116
2117         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2118                          &cached_state);
2119
2120         /* already ordered? We're done */
2121         if (PagePrivate2(page))
2122                 goto out;
2123
2124         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2125                                         PAGE_SIZE);
2126         if (ordered) {
2127                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2128                                      page_end, &cached_state, GFP_NOFS);
2129                 unlock_page(page);
2130                 btrfs_start_ordered_extent(inode, ordered, 1);
2131                 btrfs_put_ordered_extent(ordered);
2132                 goto again;
2133         }
2134
2135         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2136                                            PAGE_SIZE);
2137         if (ret) {
2138                 mapping_set_error(page->mapping, ret);
2139                 end_extent_writepage(page, ret, page_start, page_end);
2140                 ClearPageChecked(page);
2141                 goto out;
2142          }
2143
2144         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
2145                                         &cached_state, 0);
2146         if (ret) {
2147                 mapping_set_error(page->mapping, ret);
2148                 end_extent_writepage(page, ret, page_start, page_end);
2149                 ClearPageChecked(page);
2150                 goto out;
2151         }
2152
2153         ClearPageChecked(page);
2154         set_page_dirty(page);
2155 out:
2156         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2157                              &cached_state, GFP_NOFS);
2158 out_page:
2159         unlock_page(page);
2160         put_page(page);
2161         kfree(fixup);
2162         extent_changeset_free(data_reserved);
2163 }
2164
2165 /*
2166  * There are a few paths in the higher layers of the kernel that directly
2167  * set the page dirty bit without asking the filesystem if it is a
2168  * good idea.  This causes problems because we want to make sure COW
2169  * properly happens and the data=ordered rules are followed.
2170  *
2171  * In our case any range that doesn't have the ORDERED bit set
2172  * hasn't been properly setup for IO.  We kick off an async process
2173  * to fix it up.  The async helper will wait for ordered extents, set
2174  * the delalloc bit and make it safe to write the page.
2175  */
2176 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2177 {
2178         struct inode *inode = page->mapping->host;
2179         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2180         struct btrfs_writepage_fixup *fixup;
2181
2182         /* this page is properly in the ordered list */
2183         if (TestClearPagePrivate2(page))
2184                 return 0;
2185
2186         if (PageChecked(page))
2187                 return -EAGAIN;
2188
2189         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2190         if (!fixup)
2191                 return -EAGAIN;
2192
2193         SetPageChecked(page);
2194         get_page(page);
2195         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2196                         btrfs_writepage_fixup_worker, NULL, NULL);
2197         fixup->page = page;
2198         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2199         return -EBUSY;
2200 }
2201
2202 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2203                                        struct inode *inode, u64 file_pos,
2204                                        u64 disk_bytenr, u64 disk_num_bytes,
2205                                        u64 num_bytes, u64 ram_bytes,
2206                                        u8 compression, u8 encryption,
2207                                        u16 other_encoding, int extent_type)
2208 {
2209         struct btrfs_root *root = BTRFS_I(inode)->root;
2210         struct btrfs_file_extent_item *fi;
2211         struct btrfs_path *path;
2212         struct extent_buffer *leaf;
2213         struct btrfs_key ins;
2214         u64 qg_released;
2215         int extent_inserted = 0;
2216         int ret;
2217
2218         path = btrfs_alloc_path();
2219         if (!path)
2220                 return -ENOMEM;
2221
2222         /*
2223          * we may be replacing one extent in the tree with another.
2224          * The new extent is pinned in the extent map, and we don't want
2225          * to drop it from the cache until it is completely in the btree.
2226          *
2227          * So, tell btrfs_drop_extents to leave this extent in the cache.
2228          * the caller is expected to unpin it and allow it to be merged
2229          * with the others.
2230          */
2231         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2232                                    file_pos + num_bytes, NULL, 0,
2233                                    1, sizeof(*fi), &extent_inserted);
2234         if (ret)
2235                 goto out;
2236
2237         if (!extent_inserted) {
2238                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2239                 ins.offset = file_pos;
2240                 ins.type = BTRFS_EXTENT_DATA_KEY;
2241
2242                 path->leave_spinning = 1;
2243                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2244                                               sizeof(*fi));
2245                 if (ret)
2246                         goto out;
2247         }
2248         leaf = path->nodes[0];
2249         fi = btrfs_item_ptr(leaf, path->slots[0],
2250                             struct btrfs_file_extent_item);
2251         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2252         btrfs_set_file_extent_type(leaf, fi, extent_type);
2253         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2254         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2255         btrfs_set_file_extent_offset(leaf, fi, 0);
2256         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2257         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2258         btrfs_set_file_extent_compression(leaf, fi, compression);
2259         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2260         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2261
2262         btrfs_mark_buffer_dirty(leaf);
2263         btrfs_release_path(path);
2264
2265         inode_add_bytes(inode, num_bytes);
2266
2267         ins.objectid = disk_bytenr;
2268         ins.offset = disk_num_bytes;
2269         ins.type = BTRFS_EXTENT_ITEM_KEY;
2270
2271         /*
2272          * Release the reserved range from inode dirty range map, as it is
2273          * already moved into delayed_ref_head
2274          */
2275         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2276         if (ret < 0)
2277                 goto out;
2278         qg_released = ret;
2279         ret = btrfs_alloc_reserved_file_extent(trans, root->root_key.objectid,
2280                         btrfs_ino(BTRFS_I(inode)), file_pos, qg_released, &ins);
2281 out:
2282         btrfs_free_path(path);
2283
2284         return ret;
2285 }
2286
2287 /* snapshot-aware defrag */
2288 struct sa_defrag_extent_backref {
2289         struct rb_node node;
2290         struct old_sa_defrag_extent *old;
2291         u64 root_id;
2292         u64 inum;
2293         u64 file_pos;
2294         u64 extent_offset;
2295         u64 num_bytes;
2296         u64 generation;
2297 };
2298
2299 struct old_sa_defrag_extent {
2300         struct list_head list;
2301         struct new_sa_defrag_extent *new;
2302
2303         u64 extent_offset;
2304         u64 bytenr;
2305         u64 offset;
2306         u64 len;
2307         int count;
2308 };
2309
2310 struct new_sa_defrag_extent {
2311         struct rb_root root;
2312         struct list_head head;
2313         struct btrfs_path *path;
2314         struct inode *inode;
2315         u64 file_pos;
2316         u64 len;
2317         u64 bytenr;
2318         u64 disk_len;
2319         u8 compress_type;
2320 };
2321
2322 static int backref_comp(struct sa_defrag_extent_backref *b1,
2323                         struct sa_defrag_extent_backref *b2)
2324 {
2325         if (b1->root_id < b2->root_id)
2326                 return -1;
2327         else if (b1->root_id > b2->root_id)
2328                 return 1;
2329
2330         if (b1->inum < b2->inum)
2331                 return -1;
2332         else if (b1->inum > b2->inum)
2333                 return 1;
2334
2335         if (b1->file_pos < b2->file_pos)
2336                 return -1;
2337         else if (b1->file_pos > b2->file_pos)
2338                 return 1;
2339
2340         /*
2341          * [------------------------------] ===> (a range of space)
2342          *     |<--->|   |<---->| =============> (fs/file tree A)
2343          * |<---------------------------->| ===> (fs/file tree B)
2344          *
2345          * A range of space can refer to two file extents in one tree while
2346          * refer to only one file extent in another tree.
2347          *
2348          * So we may process a disk offset more than one time(two extents in A)
2349          * and locate at the same extent(one extent in B), then insert two same
2350          * backrefs(both refer to the extent in B).
2351          */
2352         return 0;
2353 }
2354
2355 static void backref_insert(struct rb_root *root,
2356                            struct sa_defrag_extent_backref *backref)
2357 {
2358         struct rb_node **p = &root->rb_node;
2359         struct rb_node *parent = NULL;
2360         struct sa_defrag_extent_backref *entry;
2361         int ret;
2362
2363         while (*p) {
2364                 parent = *p;
2365                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2366
2367                 ret = backref_comp(backref, entry);
2368                 if (ret < 0)
2369                         p = &(*p)->rb_left;
2370                 else
2371                         p = &(*p)->rb_right;
2372         }
2373
2374         rb_link_node(&backref->node, parent, p);
2375         rb_insert_color(&backref->node, root);
2376 }
2377
2378 /*
2379  * Note the backref might has changed, and in this case we just return 0.
2380  */
2381 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2382                                        void *ctx)
2383 {
2384         struct btrfs_file_extent_item *extent;
2385         struct old_sa_defrag_extent *old = ctx;
2386         struct new_sa_defrag_extent *new = old->new;
2387         struct btrfs_path *path = new->path;
2388         struct btrfs_key key;
2389         struct btrfs_root *root;
2390         struct sa_defrag_extent_backref *backref;
2391         struct extent_buffer *leaf;
2392         struct inode *inode = new->inode;
2393         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2394         int slot;
2395         int ret;
2396         u64 extent_offset;
2397         u64 num_bytes;
2398
2399         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2400             inum == btrfs_ino(BTRFS_I(inode)))
2401                 return 0;
2402
2403         key.objectid = root_id;
2404         key.type = BTRFS_ROOT_ITEM_KEY;
2405         key.offset = (u64)-1;
2406
2407         root = btrfs_read_fs_root_no_name(fs_info, &key);
2408         if (IS_ERR(root)) {
2409                 if (PTR_ERR(root) == -ENOENT)
2410                         return 0;
2411                 WARN_ON(1);
2412                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2413                          inum, offset, root_id);
2414                 return PTR_ERR(root);
2415         }
2416
2417         key.objectid = inum;
2418         key.type = BTRFS_EXTENT_DATA_KEY;
2419         if (offset > (u64)-1 << 32)
2420                 key.offset = 0;
2421         else
2422                 key.offset = offset;
2423
2424         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2425         if (WARN_ON(ret < 0))
2426                 return ret;
2427         ret = 0;
2428
2429         while (1) {
2430                 cond_resched();
2431
2432                 leaf = path->nodes[0];
2433                 slot = path->slots[0];
2434
2435                 if (slot >= btrfs_header_nritems(leaf)) {
2436                         ret = btrfs_next_leaf(root, path);
2437                         if (ret < 0) {
2438                                 goto out;
2439                         } else if (ret > 0) {
2440                                 ret = 0;
2441                                 goto out;
2442                         }
2443                         continue;
2444                 }
2445
2446                 path->slots[0]++;
2447
2448                 btrfs_item_key_to_cpu(leaf, &key, slot);
2449
2450                 if (key.objectid > inum)
2451                         goto out;
2452
2453                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2454                         continue;
2455
2456                 extent = btrfs_item_ptr(leaf, slot,
2457                                         struct btrfs_file_extent_item);
2458
2459                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2460                         continue;
2461
2462                 /*
2463                  * 'offset' refers to the exact key.offset,
2464                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2465                  * (key.offset - extent_offset).
2466                  */
2467                 if (key.offset != offset)
2468                         continue;
2469
2470                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2471                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2472
2473                 if (extent_offset >= old->extent_offset + old->offset +
2474                     old->len || extent_offset + num_bytes <=
2475                     old->extent_offset + old->offset)
2476                         continue;
2477                 break;
2478         }
2479
2480         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2481         if (!backref) {
2482                 ret = -ENOENT;
2483                 goto out;
2484         }
2485
2486         backref->root_id = root_id;
2487         backref->inum = inum;
2488         backref->file_pos = offset;
2489         backref->num_bytes = num_bytes;
2490         backref->extent_offset = extent_offset;
2491         backref->generation = btrfs_file_extent_generation(leaf, extent);
2492         backref->old = old;
2493         backref_insert(&new->root, backref);
2494         old->count++;
2495 out:
2496         btrfs_release_path(path);
2497         WARN_ON(ret);
2498         return ret;
2499 }
2500
2501 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2502                                    struct new_sa_defrag_extent *new)
2503 {
2504         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2505         struct old_sa_defrag_extent *old, *tmp;
2506         int ret;
2507
2508         new->path = path;
2509
2510         list_for_each_entry_safe(old, tmp, &new->head, list) {
2511                 ret = iterate_inodes_from_logical(old->bytenr +
2512                                                   old->extent_offset, fs_info,
2513                                                   path, record_one_backref,
2514                                                   old);
2515                 if (ret < 0 && ret != -ENOENT)
2516                         return false;
2517
2518                 /* no backref to be processed for this extent */
2519                 if (!old->count) {
2520                         list_del(&old->list);
2521                         kfree(old);
2522                 }
2523         }
2524
2525         if (list_empty(&new->head))
2526                 return false;
2527
2528         return true;
2529 }
2530
2531 static int relink_is_mergable(struct extent_buffer *leaf,
2532                               struct btrfs_file_extent_item *fi,
2533                               struct new_sa_defrag_extent *new)
2534 {
2535         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2536                 return 0;
2537
2538         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2539                 return 0;
2540
2541         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2542                 return 0;
2543
2544         if (btrfs_file_extent_encryption(leaf, fi) ||
2545             btrfs_file_extent_other_encoding(leaf, fi))
2546                 return 0;
2547
2548         return 1;
2549 }
2550
2551 /*
2552  * Note the backref might has changed, and in this case we just return 0.
2553  */
2554 static noinline int relink_extent_backref(struct btrfs_path *path,
2555                                  struct sa_defrag_extent_backref *prev,
2556                                  struct sa_defrag_extent_backref *backref)
2557 {
2558         struct btrfs_file_extent_item *extent;
2559         struct btrfs_file_extent_item *item;
2560         struct btrfs_ordered_extent *ordered;
2561         struct btrfs_trans_handle *trans;
2562         struct btrfs_root *root;
2563         struct btrfs_key key;
2564         struct extent_buffer *leaf;
2565         struct old_sa_defrag_extent *old = backref->old;
2566         struct new_sa_defrag_extent *new = old->new;
2567         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2568         struct inode *inode;
2569         struct extent_state *cached = NULL;
2570         int ret = 0;
2571         u64 start;
2572         u64 len;
2573         u64 lock_start;
2574         u64 lock_end;
2575         bool merge = false;
2576         int index;
2577
2578         if (prev && prev->root_id == backref->root_id &&
2579             prev->inum == backref->inum &&
2580             prev->file_pos + prev->num_bytes == backref->file_pos)
2581                 merge = true;
2582
2583         /* step 1: get root */
2584         key.objectid = backref->root_id;
2585         key.type = BTRFS_ROOT_ITEM_KEY;
2586         key.offset = (u64)-1;
2587
2588         index = srcu_read_lock(&fs_info->subvol_srcu);
2589
2590         root = btrfs_read_fs_root_no_name(fs_info, &key);
2591         if (IS_ERR(root)) {
2592                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2593                 if (PTR_ERR(root) == -ENOENT)
2594                         return 0;
2595                 return PTR_ERR(root);
2596         }
2597
2598         if (btrfs_root_readonly(root)) {
2599                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2600                 return 0;
2601         }
2602
2603         /* step 2: get inode */
2604         key.objectid = backref->inum;
2605         key.type = BTRFS_INODE_ITEM_KEY;
2606         key.offset = 0;
2607
2608         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2609         if (IS_ERR(inode)) {
2610                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2611                 return 0;
2612         }
2613
2614         srcu_read_unlock(&fs_info->subvol_srcu, index);
2615
2616         /* step 3: relink backref */
2617         lock_start = backref->file_pos;
2618         lock_end = backref->file_pos + backref->num_bytes - 1;
2619         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2620                          &cached);
2621
2622         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2623         if (ordered) {
2624                 btrfs_put_ordered_extent(ordered);
2625                 goto out_unlock;
2626         }
2627
2628         trans = btrfs_join_transaction(root);
2629         if (IS_ERR(trans)) {
2630                 ret = PTR_ERR(trans);
2631                 goto out_unlock;
2632         }
2633
2634         key.objectid = backref->inum;
2635         key.type = BTRFS_EXTENT_DATA_KEY;
2636         key.offset = backref->file_pos;
2637
2638         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2639         if (ret < 0) {
2640                 goto out_free_path;
2641         } else if (ret > 0) {
2642                 ret = 0;
2643                 goto out_free_path;
2644         }
2645
2646         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2647                                 struct btrfs_file_extent_item);
2648
2649         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2650             backref->generation)
2651                 goto out_free_path;
2652
2653         btrfs_release_path(path);
2654
2655         start = backref->file_pos;
2656         if (backref->extent_offset < old->extent_offset + old->offset)
2657                 start += old->extent_offset + old->offset -
2658                          backref->extent_offset;
2659
2660         len = min(backref->extent_offset + backref->num_bytes,
2661                   old->extent_offset + old->offset + old->len);
2662         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2663
2664         ret = btrfs_drop_extents(trans, root, inode, start,
2665                                  start + len, 1);
2666         if (ret)
2667                 goto out_free_path;
2668 again:
2669         key.objectid = btrfs_ino(BTRFS_I(inode));
2670         key.type = BTRFS_EXTENT_DATA_KEY;
2671         key.offset = start;
2672
2673         path->leave_spinning = 1;
2674         if (merge) {
2675                 struct btrfs_file_extent_item *fi;
2676                 u64 extent_len;
2677                 struct btrfs_key found_key;
2678
2679                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2680                 if (ret < 0)
2681                         goto out_free_path;
2682
2683                 path->slots[0]--;
2684                 leaf = path->nodes[0];
2685                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2686
2687                 fi = btrfs_item_ptr(leaf, path->slots[0],
2688                                     struct btrfs_file_extent_item);
2689                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2690
2691                 if (extent_len + found_key.offset == start &&
2692                     relink_is_mergable(leaf, fi, new)) {
2693                         btrfs_set_file_extent_num_bytes(leaf, fi,
2694                                                         extent_len + len);
2695                         btrfs_mark_buffer_dirty(leaf);
2696                         inode_add_bytes(inode, len);
2697
2698                         ret = 1;
2699                         goto out_free_path;
2700                 } else {
2701                         merge = false;
2702                         btrfs_release_path(path);
2703                         goto again;
2704                 }
2705         }
2706
2707         ret = btrfs_insert_empty_item(trans, root, path, &key,
2708                                         sizeof(*extent));
2709         if (ret) {
2710                 btrfs_abort_transaction(trans, ret);
2711                 goto out_free_path;
2712         }
2713
2714         leaf = path->nodes[0];
2715         item = btrfs_item_ptr(leaf, path->slots[0],
2716                                 struct btrfs_file_extent_item);
2717         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2718         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2719         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2720         btrfs_set_file_extent_num_bytes(leaf, item, len);
2721         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2722         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2723         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2724         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2725         btrfs_set_file_extent_encryption(leaf, item, 0);
2726         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2727
2728         btrfs_mark_buffer_dirty(leaf);
2729         inode_add_bytes(inode, len);
2730         btrfs_release_path(path);
2731
2732         ret = btrfs_inc_extent_ref(trans, fs_info, new->bytenr,
2733                         new->disk_len, 0,
2734                         backref->root_id, backref->inum,
2735                         new->file_pos); /* start - extent_offset */
2736         if (ret) {
2737                 btrfs_abort_transaction(trans, ret);
2738                 goto out_free_path;
2739         }
2740
2741         ret = 1;
2742 out_free_path:
2743         btrfs_release_path(path);
2744         path->leave_spinning = 0;
2745         btrfs_end_transaction(trans);
2746 out_unlock:
2747         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2748                              &cached, GFP_NOFS);
2749         iput(inode);
2750         return ret;
2751 }
2752
2753 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2754 {
2755         struct old_sa_defrag_extent *old, *tmp;
2756
2757         if (!new)
2758                 return;
2759
2760         list_for_each_entry_safe(old, tmp, &new->head, list) {
2761                 kfree(old);
2762         }
2763         kfree(new);
2764 }
2765
2766 static void relink_file_extents(struct new_sa_defrag_extent *new)
2767 {
2768         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2769         struct btrfs_path *path;
2770         struct sa_defrag_extent_backref *backref;
2771         struct sa_defrag_extent_backref *prev = NULL;
2772         struct inode *inode;
2773         struct btrfs_root *root;
2774         struct rb_node *node;
2775         int ret;
2776
2777         inode = new->inode;
2778         root = BTRFS_I(inode)->root;
2779
2780         path = btrfs_alloc_path();
2781         if (!path)
2782                 return;
2783
2784         if (!record_extent_backrefs(path, new)) {
2785                 btrfs_free_path(path);
2786                 goto out;
2787         }
2788         btrfs_release_path(path);
2789
2790         while (1) {
2791                 node = rb_first(&new->root);
2792                 if (!node)
2793                         break;
2794                 rb_erase(node, &new->root);
2795
2796                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2797
2798                 ret = relink_extent_backref(path, prev, backref);
2799                 WARN_ON(ret < 0);
2800
2801                 kfree(prev);
2802
2803                 if (ret == 1)
2804                         prev = backref;
2805                 else
2806                         prev = NULL;
2807                 cond_resched();
2808         }
2809         kfree(prev);
2810
2811         btrfs_free_path(path);
2812 out:
2813         free_sa_defrag_extent(new);
2814
2815         atomic_dec(&fs_info->defrag_running);
2816         wake_up(&fs_info->transaction_wait);
2817 }
2818
2819 static struct new_sa_defrag_extent *
2820 record_old_file_extents(struct inode *inode,
2821                         struct btrfs_ordered_extent *ordered)
2822 {
2823         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2824         struct btrfs_root *root = BTRFS_I(inode)->root;
2825         struct btrfs_path *path;
2826         struct btrfs_key key;
2827         struct old_sa_defrag_extent *old;
2828         struct new_sa_defrag_extent *new;
2829         int ret;
2830
2831         new = kmalloc(sizeof(*new), GFP_NOFS);
2832         if (!new)
2833                 return NULL;
2834
2835         new->inode = inode;
2836         new->file_pos = ordered->file_offset;
2837         new->len = ordered->len;
2838         new->bytenr = ordered->start;
2839         new->disk_len = ordered->disk_len;
2840         new->compress_type = ordered->compress_type;
2841         new->root = RB_ROOT;
2842         INIT_LIST_HEAD(&new->head);
2843
2844         path = btrfs_alloc_path();
2845         if (!path)
2846                 goto out_kfree;
2847
2848         key.objectid = btrfs_ino(BTRFS_I(inode));
2849         key.type = BTRFS_EXTENT_DATA_KEY;
2850         key.offset = new->file_pos;
2851
2852         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2853         if (ret < 0)
2854                 goto out_free_path;
2855         if (ret > 0 && path->slots[0] > 0)
2856                 path->slots[0]--;
2857
2858         /* find out all the old extents for the file range */
2859         while (1) {
2860                 struct btrfs_file_extent_item *extent;
2861                 struct extent_buffer *l;
2862                 int slot;
2863                 u64 num_bytes;
2864                 u64 offset;
2865                 u64 end;
2866                 u64 disk_bytenr;
2867                 u64 extent_offset;
2868
2869                 l = path->nodes[0];
2870                 slot = path->slots[0];
2871
2872                 if (slot >= btrfs_header_nritems(l)) {
2873                         ret = btrfs_next_leaf(root, path);
2874                         if (ret < 0)
2875                                 goto out_free_path;
2876                         else if (ret > 0)
2877                                 break;
2878                         continue;
2879                 }
2880
2881                 btrfs_item_key_to_cpu(l, &key, slot);
2882
2883                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2884                         break;
2885                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2886                         break;
2887                 if (key.offset >= new->file_pos + new->len)
2888                         break;
2889
2890                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2891
2892                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2893                 if (key.offset + num_bytes < new->file_pos)
2894                         goto next;
2895
2896                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2897                 if (!disk_bytenr)
2898                         goto next;
2899
2900                 extent_offset = btrfs_file_extent_offset(l, extent);
2901
2902                 old = kmalloc(sizeof(*old), GFP_NOFS);
2903                 if (!old)
2904                         goto out_free_path;
2905
2906                 offset = max(new->file_pos, key.offset);
2907                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2908
2909                 old->bytenr = disk_bytenr;
2910                 old->extent_offset = extent_offset;
2911                 old->offset = offset - key.offset;
2912                 old->len = end - offset;
2913                 old->new = new;
2914                 old->count = 0;
2915                 list_add_tail(&old->list, &new->head);
2916 next:
2917                 path->slots[0]++;
2918                 cond_resched();
2919         }
2920
2921         btrfs_free_path(path);
2922         atomic_inc(&fs_info->defrag_running);
2923
2924         return new;
2925
2926 out_free_path:
2927         btrfs_free_path(path);
2928 out_kfree:
2929         free_sa_defrag_extent(new);
2930         return NULL;
2931 }
2932
2933 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2934                                          u64 start, u64 len)
2935 {
2936         struct btrfs_block_group_cache *cache;
2937
2938         cache = btrfs_lookup_block_group(fs_info, start);
2939         ASSERT(cache);
2940
2941         spin_lock(&cache->lock);
2942         cache->delalloc_bytes -= len;
2943         spin_unlock(&cache->lock);
2944
2945         btrfs_put_block_group(cache);
2946 }
2947
2948 /* as ordered data IO finishes, this gets called so we can finish
2949  * an ordered extent if the range of bytes in the file it covers are
2950  * fully written.
2951  */
2952 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2953 {
2954         struct inode *inode = ordered_extent->inode;
2955         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2956         struct btrfs_root *root = BTRFS_I(inode)->root;
2957         struct btrfs_trans_handle *trans = NULL;
2958         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2959         struct extent_state *cached_state = NULL;
2960         struct new_sa_defrag_extent *new = NULL;
2961         int compress_type = 0;
2962         int ret = 0;
2963         u64 logical_len = ordered_extent->len;
2964         bool nolock;
2965         bool truncated = false;
2966         bool range_locked = false;
2967         bool clear_new_delalloc_bytes = false;
2968         bool clear_reserved_extent = true;
2969
2970         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2971             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2972             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2973                 clear_new_delalloc_bytes = true;
2974
2975         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2976
2977         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2978                 ret = -EIO;
2979                 goto out;
2980         }
2981
2982         btrfs_free_io_failure_record(BTRFS_I(inode),
2983                         ordered_extent->file_offset,
2984                         ordered_extent->file_offset +
2985                         ordered_extent->len - 1);
2986
2987         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2988                 truncated = true;
2989                 logical_len = ordered_extent->truncated_len;
2990                 /* Truncated the entire extent, don't bother adding */
2991                 if (!logical_len)
2992                         goto out;
2993         }
2994
2995         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2996                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2997
2998                 /*
2999                  * For mwrite(mmap + memset to write) case, we still reserve
3000                  * space for NOCOW range.
3001                  * As NOCOW won't cause a new delayed ref, just free the space
3002                  */
3003                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3004                                        ordered_extent->len);
3005                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3006                 if (nolock)
3007                         trans = btrfs_join_transaction_nolock(root);
3008                 else
3009                         trans = btrfs_join_transaction(root);
3010                 if (IS_ERR(trans)) {
3011                         ret = PTR_ERR(trans);
3012                         trans = NULL;
3013                         goto out;
3014                 }
3015                 trans->block_rsv = &fs_info->delalloc_block_rsv;
3016                 ret = btrfs_update_inode_fallback(trans, root, inode);
3017                 if (ret) /* -ENOMEM or corruption */
3018                         btrfs_abort_transaction(trans, ret);
3019                 goto out;
3020         }
3021
3022         range_locked = true;
3023         lock_extent_bits(io_tree, ordered_extent->file_offset,
3024                          ordered_extent->file_offset + ordered_extent->len - 1,
3025                          &cached_state);
3026
3027         ret = test_range_bit(io_tree, ordered_extent->file_offset,
3028                         ordered_extent->file_offset + ordered_extent->len - 1,
3029                         EXTENT_DEFRAG, 0, cached_state);
3030         if (ret) {
3031                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
3032                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
3033                         /* the inode is shared */
3034                         new = record_old_file_extents(inode, ordered_extent);
3035
3036                 clear_extent_bit(io_tree, ordered_extent->file_offset,
3037                         ordered_extent->file_offset + ordered_extent->len - 1,
3038                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
3039         }
3040
3041         if (nolock)
3042                 trans = btrfs_join_transaction_nolock(root);
3043         else
3044                 trans = btrfs_join_transaction(root);
3045         if (IS_ERR(trans)) {
3046                 ret = PTR_ERR(trans);
3047                 trans = NULL;
3048                 goto out;
3049         }
3050
3051         trans->block_rsv = &fs_info->delalloc_block_rsv;
3052
3053         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3054                 compress_type = ordered_extent->compress_type;
3055         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3056                 BUG_ON(compress_type);
3057                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3058                                        ordered_extent->len);
3059                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3060                                                 ordered_extent->file_offset,
3061                                                 ordered_extent->file_offset +
3062                                                 logical_len);
3063         } else {
3064                 BUG_ON(root == fs_info->tree_root);
3065                 ret = insert_reserved_file_extent(trans, inode,
3066                                                 ordered_extent->file_offset,
3067                                                 ordered_extent->start,
3068                                                 ordered_extent->disk_len,
3069                                                 logical_len, logical_len,
3070                                                 compress_type, 0, 0,
3071                                                 BTRFS_FILE_EXTENT_REG);
3072                 if (!ret) {
3073                         clear_reserved_extent = false;
3074                         btrfs_release_delalloc_bytes(fs_info,
3075                                                      ordered_extent->start,
3076                                                      ordered_extent->disk_len);
3077                 }
3078         }
3079         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3080                            ordered_extent->file_offset, ordered_extent->len,
3081                            trans->transid);
3082         if (ret < 0) {
3083                 btrfs_abort_transaction(trans, ret);
3084                 goto out;
3085         }
3086
3087         add_pending_csums(trans, inode, &ordered_extent->list);
3088
3089         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3090         ret = btrfs_update_inode_fallback(trans, root, inode);
3091         if (ret) { /* -ENOMEM or corruption */
3092                 btrfs_abort_transaction(trans, ret);
3093                 goto out;
3094         }
3095         ret = 0;
3096 out:
3097         if (range_locked || clear_new_delalloc_bytes) {
3098                 unsigned int clear_bits = 0;
3099
3100                 if (range_locked)
3101                         clear_bits |= EXTENT_LOCKED;
3102                 if (clear_new_delalloc_bytes)
3103                         clear_bits |= EXTENT_DELALLOC_NEW;
3104                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3105                                  ordered_extent->file_offset,
3106                                  ordered_extent->file_offset +
3107                                  ordered_extent->len - 1,
3108                                  clear_bits,
3109                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3110                                  0, &cached_state, GFP_NOFS);
3111         }
3112
3113         if (root != fs_info->tree_root)
3114                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3115                                 ordered_extent->len);
3116         if (trans)
3117                 btrfs_end_transaction(trans);
3118
3119         if (ret || truncated) {
3120                 u64 start, end;
3121
3122                 if (truncated)
3123                         start = ordered_extent->file_offset + logical_len;
3124                 else
3125                         start = ordered_extent->file_offset;
3126                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3127                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
3128
3129                 /* Drop the cache for the part of the extent we didn't write. */
3130                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3131
3132                 /*
3133                  * If the ordered extent had an IOERR or something else went
3134                  * wrong we need to return the space for this ordered extent
3135                  * back to the allocator.  We only free the extent in the
3136                  * truncated case if we didn't write out the extent at all.
3137                  *
3138                  * If we made it past insert_reserved_file_extent before we
3139                  * errored out then we don't need to do this as the accounting
3140                  * has already been done.
3141                  */
3142                 if ((ret || !logical_len) &&
3143                     clear_reserved_extent &&
3144                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3145                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3146                         btrfs_free_reserved_extent(fs_info,
3147                                                    ordered_extent->start,
3148                                                    ordered_extent->disk_len, 1);
3149         }
3150
3151
3152         /*
3153          * This needs to be done to make sure anybody waiting knows we are done
3154          * updating everything for this ordered extent.
3155          */
3156         btrfs_remove_ordered_extent(inode, ordered_extent);
3157
3158         /* for snapshot-aware defrag */
3159         if (new) {
3160                 if (ret) {
3161                         free_sa_defrag_extent(new);
3162                         atomic_dec(&fs_info->defrag_running);
3163                 } else {
3164                         relink_file_extents(new);
3165                 }
3166         }
3167
3168         /* once for us */
3169         btrfs_put_ordered_extent(ordered_extent);
3170         /* once for the tree */
3171         btrfs_put_ordered_extent(ordered_extent);
3172
3173         /* Try to release some metadata so we don't get an OOM but don't wait */
3174         btrfs_btree_balance_dirty_nodelay(fs_info);
3175
3176         return ret;
3177 }
3178
3179 static void finish_ordered_fn(struct btrfs_work *work)
3180 {
3181         struct btrfs_ordered_extent *ordered_extent;
3182         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3183         btrfs_finish_ordered_io(ordered_extent);
3184 }
3185
3186 static void btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3187                                 struct extent_state *state, int uptodate)
3188 {
3189         struct inode *inode = page->mapping->host;
3190         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3191         struct btrfs_ordered_extent *ordered_extent = NULL;
3192         struct btrfs_workqueue *wq;
3193         btrfs_work_func_t func;
3194
3195         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3196
3197         ClearPagePrivate2(page);
3198         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3199                                             end - start + 1, uptodate))
3200                 return;
3201
3202         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
3203                 wq = fs_info->endio_freespace_worker;
3204                 func = btrfs_freespace_write_helper;
3205         } else {
3206                 wq = fs_info->endio_write_workers;
3207                 func = btrfs_endio_write_helper;
3208         }
3209
3210         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3211                         NULL);
3212         btrfs_queue_work(wq, &ordered_extent->work);
3213 }
3214
3215 static int __readpage_endio_check(struct inode *inode,
3216                                   struct btrfs_io_bio *io_bio,
3217                                   int icsum, struct page *page,
3218                                   int pgoff, u64 start, size_t len)
3219 {
3220         char *kaddr;
3221         u32 csum_expected;
3222         u32 csum = ~(u32)0;
3223
3224         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3225
3226         kaddr = kmap_atomic(page);
3227         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3228         btrfs_csum_final(csum, (u8 *)&csum);
3229         if (csum != csum_expected)
3230                 goto zeroit;
3231
3232         kunmap_atomic(kaddr);
3233         return 0;
3234 zeroit:
3235         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3236                                     io_bio->mirror_num);
3237         memset(kaddr + pgoff, 1, len);
3238         flush_dcache_page(page);
3239         kunmap_atomic(kaddr);
3240         return -EIO;
3241 }
3242
3243 /*
3244  * when reads are done, we need to check csums to verify the data is correct
3245  * if there's a match, we allow the bio to finish.  If not, the code in
3246  * extent_io.c will try to find good copies for us.
3247  */
3248 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3249                                       u64 phy_offset, struct page *page,
3250                                       u64 start, u64 end, int mirror)
3251 {
3252         size_t offset = start - page_offset(page);
3253         struct inode *inode = page->mapping->host;
3254         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3255         struct btrfs_root *root = BTRFS_I(inode)->root;
3256
3257         if (PageChecked(page)) {
3258                 ClearPageChecked(page);
3259                 return 0;
3260         }
3261
3262         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3263                 return 0;
3264
3265         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3266             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3267                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3268                 return 0;
3269         }
3270
3271         phy_offset >>= inode->i_sb->s_blocksize_bits;
3272         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3273                                       start, (size_t)(end - start + 1));
3274 }
3275
3276 void btrfs_add_delayed_iput(struct inode *inode)
3277 {
3278         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3279         struct btrfs_inode *binode = BTRFS_I(inode);
3280
3281         if (atomic_add_unless(&inode->i_count, -1, 1))
3282                 return;
3283
3284         spin_lock(&fs_info->delayed_iput_lock);
3285         if (binode->delayed_iput_count == 0) {
3286                 ASSERT(list_empty(&binode->delayed_iput));
3287                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3288         } else {
3289                 binode->delayed_iput_count++;
3290         }
3291         spin_unlock(&fs_info->delayed_iput_lock);
3292 }
3293
3294 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3295 {
3296
3297         spin_lock(&fs_info->delayed_iput_lock);
3298         while (!list_empty(&fs_info->delayed_iputs)) {
3299                 struct btrfs_inode *inode;
3300
3301                 inode = list_first_entry(&fs_info->delayed_iputs,
3302                                 struct btrfs_inode, delayed_iput);
3303                 if (inode->delayed_iput_count) {
3304                         inode->delayed_iput_count--;
3305                         list_move_tail(&inode->delayed_iput,
3306                                         &fs_info->delayed_iputs);
3307                 } else {
3308                         list_del_init(&inode->delayed_iput);
3309                 }
3310                 spin_unlock(&fs_info->delayed_iput_lock);
3311                 iput(&inode->vfs_inode);
3312                 spin_lock(&fs_info->delayed_iput_lock);
3313         }
3314         spin_unlock(&fs_info->delayed_iput_lock);
3315 }
3316
3317 /*
3318  * This is called in transaction commit time. If there are no orphan
3319  * files in the subvolume, it removes orphan item and frees block_rsv
3320  * structure.
3321  */
3322 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3323                               struct btrfs_root *root)
3324 {
3325         struct btrfs_fs_info *fs_info = root->fs_info;
3326         struct btrfs_block_rsv *block_rsv;
3327         int ret;
3328
3329         if (atomic_read(&root->orphan_inodes) ||
3330             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3331                 return;
3332
3333         spin_lock(&root->orphan_lock);
3334         if (atomic_read(&root->orphan_inodes)) {
3335                 spin_unlock(&root->orphan_lock);
3336                 return;
3337         }
3338
3339         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3340                 spin_unlock(&root->orphan_lock);
3341                 return;
3342         }
3343
3344         block_rsv = root->orphan_block_rsv;
3345         root->orphan_block_rsv = NULL;
3346         spin_unlock(&root->orphan_lock);
3347
3348         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3349             btrfs_root_refs(&root->root_item) > 0) {
3350                 ret = btrfs_del_orphan_item(trans, fs_info->tree_root,
3351                                             root->root_key.objectid);
3352                 if (ret)
3353                         btrfs_abort_transaction(trans, ret);
3354                 else
3355                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3356                                   &root->state);
3357         }
3358
3359         if (block_rsv) {
3360                 WARN_ON(block_rsv->size > 0);
3361                 btrfs_free_block_rsv(fs_info, block_rsv);
3362         }
3363 }
3364
3365 /*
3366  * This creates an orphan entry for the given inode in case something goes
3367  * wrong in the middle of an unlink/truncate.
3368  *
3369  * NOTE: caller of this function should reserve 5 units of metadata for
3370  *       this function.
3371  */
3372 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3373                 struct btrfs_inode *inode)
3374 {
3375         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
3376         struct btrfs_root *root = inode->root;
3377         struct btrfs_block_rsv *block_rsv = NULL;
3378         int reserve = 0;
3379         int insert = 0;
3380         int ret;
3381
3382         if (!root->orphan_block_rsv) {
3383                 block_rsv = btrfs_alloc_block_rsv(fs_info,
3384                                                   BTRFS_BLOCK_RSV_TEMP);
3385                 if (!block_rsv)
3386                         return -ENOMEM;
3387         }
3388
3389         spin_lock(&root->orphan_lock);
3390         if (!root->orphan_block_rsv) {
3391                 root->orphan_block_rsv = block_rsv;
3392         } else if (block_rsv) {
3393                 btrfs_free_block_rsv(fs_info, block_rsv);
3394                 block_rsv = NULL;
3395         }
3396
3397         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3398                               &inode->runtime_flags)) {
3399 #if 0
3400                 /*
3401                  * For proper ENOSPC handling, we should do orphan
3402                  * cleanup when mounting. But this introduces backward
3403                  * compatibility issue.
3404                  */
3405                 if (!xchg(&root->orphan_item_inserted, 1))
3406                         insert = 2;
3407                 else
3408                         insert = 1;
3409 #endif
3410                 insert = 1;
3411                 atomic_inc(&root->orphan_inodes);
3412         }
3413
3414         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3415                               &inode->runtime_flags))
3416                 reserve = 1;
3417         spin_unlock(&root->orphan_lock);
3418
3419         /* grab metadata reservation from transaction handle */
3420         if (reserve) {
3421                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3422                 ASSERT(!ret);
3423                 if (ret) {
3424                         /*
3425                          * dec doesn't need spin_lock as ->orphan_block_rsv
3426                          * would be released only if ->orphan_inodes is
3427                          * zero.
3428                          */
3429                         atomic_dec(&root->orphan_inodes);
3430                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3431                                   &inode->runtime_flags);
3432                         if (insert)
3433                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3434                                           &inode->runtime_flags);
3435                         return ret;
3436                 }
3437         }
3438
3439         /* insert an orphan item to track this unlinked/truncated file */
3440         if (insert >= 1) {
3441                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3442                 if (ret) {
3443                         if (reserve) {
3444                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3445                                           &inode->runtime_flags);
3446                                 btrfs_orphan_release_metadata(inode);
3447                         }
3448                         /*
3449                          * btrfs_orphan_commit_root may race with us and set
3450                          * ->orphan_block_rsv to zero, in order to avoid that,
3451                          * decrease ->orphan_inodes after everything is done.
3452                          */
3453                         atomic_dec(&root->orphan_inodes);
3454                         if (ret != -EEXIST) {
3455                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3456                                           &inode->runtime_flags);
3457                                 btrfs_abort_transaction(trans, ret);
3458                                 return ret;
3459                         }
3460                 }
3461                 ret = 0;
3462         }
3463
3464         /* insert an orphan item to track subvolume contains orphan files */
3465         if (insert >= 2) {
3466                 ret = btrfs_insert_orphan_item(trans, fs_info->tree_root,
3467                                                root->root_key.objectid);
3468                 if (ret && ret != -EEXIST) {
3469                         btrfs_abort_transaction(trans, ret);
3470                         return ret;
3471                 }
3472         }
3473         return 0;
3474 }
3475
3476 /*
3477  * We have done the truncate/delete so we can go ahead and remove the orphan
3478  * item for this particular inode.
3479  */
3480 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3481                             struct btrfs_inode *inode)
3482 {
3483         struct btrfs_root *root = inode->root;
3484         int delete_item = 0;
3485         int ret = 0;
3486
3487         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3488                                &inode->runtime_flags))
3489                 delete_item = 1;
3490
3491         if (delete_item && trans)
3492                 ret = btrfs_del_orphan_item(trans, root, btrfs_ino(inode));
3493
3494         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3495                                &inode->runtime_flags))
3496                 btrfs_orphan_release_metadata(inode);
3497
3498         /*
3499          * btrfs_orphan_commit_root may race with us and set ->orphan_block_rsv
3500          * to zero, in order to avoid that, decrease ->orphan_inodes after
3501          * everything is done.
3502          */
3503         if (delete_item)
3504                 atomic_dec(&root->orphan_inodes);
3505
3506         return ret;
3507 }
3508
3509 /*
3510  * this cleans up any orphans that may be left on the list from the last use
3511  * of this root.
3512  */
3513 int btrfs_orphan_cleanup(struct btrfs_root *root)
3514 {
3515         struct btrfs_fs_info *fs_info = root->fs_info;
3516         struct btrfs_path *path;
3517         struct extent_buffer *leaf;
3518         struct btrfs_key key, found_key;
3519         struct btrfs_trans_handle *trans;
3520         struct inode *inode;
3521         u64 last_objectid = 0;
3522         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3523
3524         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3525                 return 0;
3526
3527         path = btrfs_alloc_path();
3528         if (!path) {
3529                 ret = -ENOMEM;
3530                 goto out;
3531         }
3532         path->reada = READA_BACK;
3533
3534         key.objectid = BTRFS_ORPHAN_OBJECTID;
3535         key.type = BTRFS_ORPHAN_ITEM_KEY;
3536         key.offset = (u64)-1;
3537
3538         while (1) {
3539                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3540                 if (ret < 0)
3541                         goto out;
3542
3543                 /*
3544                  * if ret == 0 means we found what we were searching for, which
3545                  * is weird, but possible, so only screw with path if we didn't
3546                  * find the key and see if we have stuff that matches
3547                  */
3548                 if (ret > 0) {
3549                         ret = 0;
3550                         if (path->slots[0] == 0)
3551                                 break;
3552                         path->slots[0]--;
3553                 }
3554
3555                 /* pull out the item */
3556                 leaf = path->nodes[0];
3557                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3558
3559                 /* make sure the item matches what we want */
3560                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3561                         break;
3562                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3563                         break;
3564
3565                 /* release the path since we're done with it */
3566                 btrfs_release_path(path);
3567
3568                 /*
3569                  * this is where we are basically btrfs_lookup, without the
3570                  * crossing root thing.  we store the inode number in the
3571                  * offset of the orphan item.
3572                  */
3573
3574                 if (found_key.offset == last_objectid) {
3575                         btrfs_err(fs_info,
3576                                   "Error removing orphan entry, stopping orphan cleanup");
3577                         ret = -EINVAL;
3578                         goto out;
3579                 }
3580
3581                 last_objectid = found_key.offset;
3582
3583                 found_key.objectid = found_key.offset;
3584                 found_key.type = BTRFS_INODE_ITEM_KEY;
3585                 found_key.offset = 0;
3586                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3587                 ret = PTR_ERR_OR_ZERO(inode);
3588                 if (ret && ret != -ENOENT)
3589                         goto out;
3590
3591                 if (ret == -ENOENT && root == fs_info->tree_root) {
3592                         struct btrfs_root *dead_root;
3593                         struct btrfs_fs_info *fs_info = root->fs_info;
3594                         int is_dead_root = 0;
3595
3596                         /*
3597                          * this is an orphan in the tree root. Currently these
3598                          * could come from 2 sources:
3599                          *  a) a snapshot deletion in progress
3600                          *  b) a free space cache inode
3601                          * We need to distinguish those two, as the snapshot
3602                          * orphan must not get deleted.
3603                          * find_dead_roots already ran before us, so if this
3604                          * is a snapshot deletion, we should find the root
3605                          * in the dead_roots list
3606                          */
3607                         spin_lock(&fs_info->trans_lock);
3608                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3609                                             root_list) {
3610                                 if (dead_root->root_key.objectid ==
3611                                     found_key.objectid) {
3612                                         is_dead_root = 1;
3613                                         break;
3614                                 }
3615                         }
3616                         spin_unlock(&fs_info->trans_lock);
3617                         if (is_dead_root) {
3618                                 /* prevent this orphan from being found again */
3619                                 key.offset = found_key.objectid - 1;
3620                                 continue;
3621                         }
3622                 }
3623                 /*
3624                  * Inode is already gone but the orphan item is still there,
3625                  * kill the orphan item.
3626                  */
3627                 if (ret == -ENOENT) {
3628                         trans = btrfs_start_transaction(root, 1);
3629                         if (IS_ERR(trans)) {
3630                                 ret = PTR_ERR(trans);
3631                                 goto out;
3632                         }
3633                         btrfs_debug(fs_info, "auto deleting %Lu",
3634                                     found_key.objectid);
3635                         ret = btrfs_del_orphan_item(trans, root,
3636                                                     found_key.objectid);
3637                         btrfs_end_transaction(trans);
3638                         if (ret)
3639                                 goto out;
3640                         continue;
3641                 }
3642
3643                 /*
3644                  * add this inode to the orphan list so btrfs_orphan_del does
3645                  * the proper thing when we hit it
3646                  */
3647                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3648                         &BTRFS_I(inode)->runtime_flags);
3649                 atomic_inc(&root->orphan_inodes);
3650
3651                 /* if we have links, this was a truncate, lets do that */
3652                 if (inode->i_nlink) {
3653                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3654                                 iput(inode);
3655                                 continue;
3656                         }
3657                         nr_truncate++;
3658
3659                         /* 1 for the orphan item deletion. */
3660                         trans = btrfs_start_transaction(root, 1);
3661                         if (IS_ERR(trans)) {
3662                                 iput(inode);
3663                                 ret = PTR_ERR(trans);
3664                                 goto out;
3665                         }
3666                         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
3667                         btrfs_end_transaction(trans);
3668                         if (ret) {
3669                                 iput(inode);
3670                                 goto out;
3671                         }
3672
3673                         ret = btrfs_truncate(inode);
3674                         if (ret)
3675                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
3676                 } else {
3677                         nr_unlink++;
3678                 }
3679
3680                 /* this will do delete_inode and everything for us */
3681                 iput(inode);
3682                 if (ret)
3683                         goto out;
3684         }
3685         /* release the path since we're done with it */
3686         btrfs_release_path(path);
3687
3688         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3689
3690         if (root->orphan_block_rsv)
3691                 btrfs_block_rsv_release(fs_info, root->orphan_block_rsv,
3692                                         (u64)-1);
3693
3694         if (root->orphan_block_rsv ||
3695             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3696                 trans = btrfs_join_transaction(root);
3697                 if (!IS_ERR(trans))
3698                         btrfs_end_transaction(trans);
3699         }
3700
3701         if (nr_unlink)
3702                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3703         if (nr_truncate)
3704                 btrfs_debug(fs_info, "truncated %d orphans", nr_truncate);
3705
3706 out:
3707         if (ret)
3708                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3709         btrfs_free_path(path);
3710         return ret;
3711 }
3712
3713 /*
3714  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3715  * don't find any xattrs, we know there can't be any acls.
3716  *
3717  * slot is the slot the inode is in, objectid is the objectid of the inode
3718  */
3719 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3720                                           int slot, u64 objectid,
3721                                           int *first_xattr_slot)
3722 {
3723         u32 nritems = btrfs_header_nritems(leaf);
3724         struct btrfs_key found_key;
3725         static u64 xattr_access = 0;
3726         static u64 xattr_default = 0;
3727         int scanned = 0;
3728
3729         if (!xattr_access) {
3730                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3731                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3732                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3733                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3734         }
3735
3736         slot++;
3737         *first_xattr_slot = -1;
3738         while (slot < nritems) {
3739                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3740
3741                 /* we found a different objectid, there must not be acls */
3742                 if (found_key.objectid != objectid)
3743                         return 0;
3744
3745                 /* we found an xattr, assume we've got an acl */
3746                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3747                         if (*first_xattr_slot == -1)
3748                                 *first_xattr_slot = slot;
3749                         if (found_key.offset == xattr_access ||
3750                             found_key.offset == xattr_default)
3751                                 return 1;
3752                 }
3753
3754                 /*
3755                  * we found a key greater than an xattr key, there can't
3756                  * be any acls later on
3757                  */
3758                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3759                         return 0;
3760
3761                 slot++;
3762                 scanned++;
3763
3764                 /*
3765                  * it goes inode, inode backrefs, xattrs, extents,
3766                  * so if there are a ton of hard links to an inode there can
3767                  * be a lot of backrefs.  Don't waste time searching too hard,
3768                  * this is just an optimization
3769                  */
3770                 if (scanned >= 8)
3771                         break;
3772         }
3773         /* we hit the end of the leaf before we found an xattr or
3774          * something larger than an xattr.  We have to assume the inode
3775          * has acls
3776          */
3777         if (*first_xattr_slot == -1)
3778                 *first_xattr_slot = slot;
3779         return 1;
3780 }
3781
3782 /*
3783  * read an inode from the btree into the in-memory inode
3784  */
3785 static int btrfs_read_locked_inode(struct inode *inode)
3786 {
3787         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3788         struct btrfs_path *path;
3789         struct extent_buffer *leaf;
3790         struct btrfs_inode_item *inode_item;
3791         struct btrfs_root *root = BTRFS_I(inode)->root;
3792         struct btrfs_key location;
3793         unsigned long ptr;
3794         int maybe_acls;
3795         u32 rdev;
3796         int ret;
3797         bool filled = false;
3798         int first_xattr_slot;
3799
3800         ret = btrfs_fill_inode(inode, &rdev);
3801         if (!ret)
3802                 filled = true;
3803
3804         path = btrfs_alloc_path();
3805         if (!path) {
3806                 ret = -ENOMEM;
3807                 goto make_bad;
3808         }
3809
3810         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3811
3812         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3813         if (ret) {
3814                 if (ret > 0)
3815                         ret = -ENOENT;
3816                 goto make_bad;
3817         }
3818
3819         leaf = path->nodes[0];
3820
3821         if (filled)
3822                 goto cache_index;
3823
3824         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3825                                     struct btrfs_inode_item);
3826         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3827         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3828         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3829         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3830         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3831
3832         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3833         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3834
3835         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3836         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3837
3838         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3839         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3840
3841         BTRFS_I(inode)->i_otime.tv_sec =
3842                 btrfs_timespec_sec(leaf, &inode_item->otime);
3843         BTRFS_I(inode)->i_otime.tv_nsec =
3844                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3845
3846         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3847         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3848         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3849
3850         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3851         inode->i_generation = BTRFS_I(inode)->generation;
3852         inode->i_rdev = 0;
3853         rdev = btrfs_inode_rdev(leaf, inode_item);
3854
3855         BTRFS_I(inode)->index_cnt = (u64)-1;
3856         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3857
3858 cache_index:
3859         /*
3860          * If we were modified in the current generation and evicted from memory
3861          * and then re-read we need to do a full sync since we don't have any
3862          * idea about which extents were modified before we were evicted from
3863          * cache.
3864          *
3865          * This is required for both inode re-read from disk and delayed inode
3866          * in delayed_nodes_tree.
3867          */
3868         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3869                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3870                         &BTRFS_I(inode)->runtime_flags);
3871
3872         /*
3873          * We don't persist the id of the transaction where an unlink operation
3874          * against the inode was last made. So here we assume the inode might
3875          * have been evicted, and therefore the exact value of last_unlink_trans
3876          * lost, and set it to last_trans to avoid metadata inconsistencies
3877          * between the inode and its parent if the inode is fsync'ed and the log
3878          * replayed. For example, in the scenario:
3879          *
3880          * touch mydir/foo
3881          * ln mydir/foo mydir/bar
3882          * sync
3883          * unlink mydir/bar
3884          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3885          * xfs_io -c fsync mydir/foo
3886          * <power failure>
3887          * mount fs, triggers fsync log replay
3888          *
3889          * We must make sure that when we fsync our inode foo we also log its
3890          * parent inode, otherwise after log replay the parent still has the
3891          * dentry with the "bar" name but our inode foo has a link count of 1
3892          * and doesn't have an inode ref with the name "bar" anymore.
3893          *
3894          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3895          * but it guarantees correctness at the expense of occasional full
3896          * transaction commits on fsync if our inode is a directory, or if our
3897          * inode is not a directory, logging its parent unnecessarily.
3898          */
3899         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3900
3901         path->slots[0]++;
3902         if (inode->i_nlink != 1 ||
3903             path->slots[0] >= btrfs_header_nritems(leaf))
3904                 goto cache_acl;
3905
3906         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3907         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3908                 goto cache_acl;
3909
3910         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3911         if (location.type == BTRFS_INODE_REF_KEY) {
3912                 struct btrfs_inode_ref *ref;
3913
3914                 ref = (struct btrfs_inode_ref *)ptr;
3915                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3916         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3917                 struct btrfs_inode_extref *extref;
3918
3919                 extref = (struct btrfs_inode_extref *)ptr;
3920                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3921                                                                      extref);
3922         }
3923 cache_acl:
3924         /*
3925          * try to precache a NULL acl entry for files that don't have
3926          * any xattrs or acls
3927          */
3928         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3929                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3930         if (first_xattr_slot != -1) {
3931                 path->slots[0] = first_xattr_slot;
3932                 ret = btrfs_load_inode_props(inode, path);
3933                 if (ret)
3934                         btrfs_err(fs_info,
3935                                   "error loading props for ino %llu (root %llu): %d",
3936                                   btrfs_ino(BTRFS_I(inode)),
3937                                   root->root_key.objectid, ret);
3938         }
3939         btrfs_free_path(path);
3940
3941         if (!maybe_acls)
3942                 cache_no_acl(inode);
3943
3944         switch (inode->i_mode & S_IFMT) {
3945         case S_IFREG:
3946                 inode->i_mapping->a_ops = &btrfs_aops;
3947                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3948                 inode->i_fop = &btrfs_file_operations;
3949                 inode->i_op = &btrfs_file_inode_operations;
3950                 break;
3951         case S_IFDIR:
3952                 inode->i_fop = &btrfs_dir_file_operations;
3953                 inode->i_op = &btrfs_dir_inode_operations;
3954                 break;
3955         case S_IFLNK:
3956                 inode->i_op = &btrfs_symlink_inode_operations;
3957                 inode_nohighmem(inode);
3958                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3959                 break;
3960         default:
3961                 inode->i_op = &btrfs_special_inode_operations;
3962                 init_special_inode(inode, inode->i_mode, rdev);
3963                 break;
3964         }
3965
3966         btrfs_update_iflags(inode);
3967         return 0;
3968
3969 make_bad:
3970         btrfs_free_path(path);
3971         make_bad_inode(inode);
3972         return ret;
3973 }
3974
3975 /*
3976  * given a leaf and an inode, copy the inode fields into the leaf
3977  */
3978 static void fill_inode_item(struct btrfs_trans_handle *trans,
3979                             struct extent_buffer *leaf,
3980                             struct btrfs_inode_item *item,
3981                             struct inode *inode)
3982 {
3983         struct btrfs_map_token token;
3984
3985         btrfs_init_map_token(&token);
3986
3987         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3988         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3989         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3990                                    &token);
3991         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3992         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3993
3994         btrfs_set_token_timespec_sec(leaf, &item->atime,
3995                                      inode->i_atime.tv_sec, &token);
3996         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3997                                       inode->i_atime.tv_nsec, &token);
3998
3999         btrfs_set_token_timespec_sec(leaf, &item->mtime,
4000                                      inode->i_mtime.tv_sec, &token);
4001         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
4002                                       inode->i_mtime.tv_nsec, &token);
4003
4004         btrfs_set_token_timespec_sec(leaf, &item->ctime,
4005                                      inode->i_ctime.tv_sec, &token);
4006         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
4007                                       inode->i_ctime.tv_nsec, &token);
4008
4009         btrfs_set_token_timespec_sec(leaf, &item->otime,
4010                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
4011         btrfs_set_token_timespec_nsec(leaf, &item->otime,
4012                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
4013
4014         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
4015                                      &token);
4016         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
4017                                          &token);
4018         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
4019         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
4020         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
4021         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
4022         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
4023 }
4024
4025 /*
4026  * copy everything in the in-memory inode into the btree.
4027  */
4028 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
4029                                 struct btrfs_root *root, struct inode *inode)
4030 {
4031         struct btrfs_inode_item *inode_item;
4032         struct btrfs_path *path;
4033         struct extent_buffer *leaf;
4034         int ret;
4035
4036         path = btrfs_alloc_path();
4037         if (!path)
4038                 return -ENOMEM;
4039
4040         path->leave_spinning = 1;
4041         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
4042                                  1);
4043         if (ret) {
4044                 if (ret > 0)
4045                         ret = -ENOENT;
4046                 goto failed;
4047         }
4048
4049         leaf = path->nodes[0];
4050         inode_item = btrfs_item_ptr(leaf, path->slots[0],
4051                                     struct btrfs_inode_item);
4052
4053         fill_inode_item(trans, leaf, inode_item, inode);
4054         btrfs_mark_buffer_dirty(leaf);
4055         btrfs_set_inode_last_trans(trans, inode);
4056         ret = 0;
4057 failed:
4058         btrfs_free_path(path);
4059         return ret;
4060 }
4061
4062 /*
4063  * copy everything in the in-memory inode into the btree.
4064  */
4065 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
4066                                 struct btrfs_root *root, struct inode *inode)
4067 {
4068         struct btrfs_fs_info *fs_info = root->fs_info;
4069         int ret;
4070
4071         /*
4072          * If the inode is a free space inode, we can deadlock during commit
4073          * if we put it into the delayed code.
4074          *
4075          * The data relocation inode should also be directly updated
4076          * without delay
4077          */
4078         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
4079             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
4080             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
4081                 btrfs_update_root_times(trans, root);
4082
4083                 ret = btrfs_delayed_update_inode(trans, root, inode);
4084                 if (!ret)
4085                         btrfs_set_inode_last_trans(trans, inode);
4086                 return ret;
4087         }
4088
4089         return btrfs_update_inode_item(trans, root, inode);
4090 }
4091
4092 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
4093                                          struct btrfs_root *root,
4094                                          struct inode *inode)
4095 {
4096         int ret;
4097
4098         ret = btrfs_update_inode(trans, root, inode);
4099         if (ret == -ENOSPC)
4100                 return btrfs_update_inode_item(trans, root, inode);
4101         return ret;
4102 }
4103
4104 /*
4105  * unlink helper that gets used here in inode.c and in the tree logging
4106  * recovery code.  It remove a link in a directory with a given name, and
4107  * also drops the back refs in the inode to the directory
4108  */
4109 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4110                                 struct btrfs_root *root,
4111                                 struct btrfs_inode *dir,
4112                                 struct btrfs_inode *inode,
4113                                 const char *name, int name_len)
4114 {
4115         struct btrfs_fs_info *fs_info = root->fs_info;
4116         struct btrfs_path *path;
4117         int ret = 0;
4118         struct extent_buffer *leaf;
4119         struct btrfs_dir_item *di;
4120         struct btrfs_key key;
4121         u64 index;
4122         u64 ino = btrfs_ino(inode);
4123         u64 dir_ino = btrfs_ino(dir);
4124
4125         path = btrfs_alloc_path();
4126         if (!path) {
4127                 ret = -ENOMEM;
4128                 goto out;
4129         }
4130
4131         path->leave_spinning = 1;
4132         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4133                                     name, name_len, -1);
4134         if (IS_ERR(di)) {
4135                 ret = PTR_ERR(di);
4136                 goto err;
4137         }
4138         if (!di) {
4139                 ret = -ENOENT;
4140                 goto err;
4141         }
4142         leaf = path->nodes[0];
4143         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4144         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4145         if (ret)
4146                 goto err;
4147         btrfs_release_path(path);
4148
4149         /*
4150          * If we don't have dir index, we have to get it by looking up
4151          * the inode ref, since we get the inode ref, remove it directly,
4152          * it is unnecessary to do delayed deletion.
4153          *
4154          * But if we have dir index, needn't search inode ref to get it.
4155          * Since the inode ref is close to the inode item, it is better
4156          * that we delay to delete it, and just do this deletion when
4157          * we update the inode item.
4158          */
4159         if (inode->dir_index) {
4160                 ret = btrfs_delayed_delete_inode_ref(inode);
4161                 if (!ret) {
4162                         index = inode->dir_index;
4163                         goto skip_backref;
4164                 }
4165         }
4166
4167         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4168                                   dir_ino, &index);
4169         if (ret) {
4170                 btrfs_info(fs_info,
4171                         "failed to delete reference to %.*s, inode %llu parent %llu",
4172                         name_len, name, ino, dir_ino);
4173                 btrfs_abort_transaction(trans, ret);
4174                 goto err;
4175         }
4176 skip_backref:
4177         ret = btrfs_delete_delayed_dir_index(trans, fs_info, dir, index);
4178         if (ret) {
4179                 btrfs_abort_transaction(trans, ret);
4180                 goto err;
4181         }
4182
4183         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
4184                         dir_ino);
4185         if (ret != 0 && ret != -ENOENT) {
4186                 btrfs_abort_transaction(trans, ret);
4187                 goto err;
4188         }
4189
4190         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
4191                         index);
4192         if (ret == -ENOENT)
4193                 ret = 0;
4194         else if (ret)
4195                 btrfs_abort_transaction(trans, ret);
4196 err:
4197         btrfs_free_path(path);
4198         if (ret)
4199                 goto out;
4200
4201         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
4202         inode_inc_iversion(&inode->vfs_inode);
4203         inode_inc_iversion(&dir->vfs_inode);
4204         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4205                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4206         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4207 out:
4208         return ret;
4209 }
4210
4211 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4212                        struct btrfs_root *root,
4213                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4214                        const char *name, int name_len)
4215 {
4216         int ret;
4217         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4218         if (!ret) {
4219                 drop_nlink(&inode->vfs_inode);
4220                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4221         }
4222         return ret;
4223 }
4224
4225 /*
4226  * helper to start transaction for unlink and rmdir.
4227  *
4228  * unlink and rmdir are special in btrfs, they do not always free space, so
4229  * if we cannot make our reservations the normal way try and see if there is
4230  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4231  * allow the unlink to occur.
4232  */
4233 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4234 {
4235         struct btrfs_root *root = BTRFS_I(dir)->root;
4236
4237         /*
4238          * 1 for the possible orphan item
4239          * 1 for the dir item
4240          * 1 for the dir index
4241          * 1 for the inode ref
4242          * 1 for the inode
4243          */
4244         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4245 }
4246
4247 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4248 {
4249         struct btrfs_root *root = BTRFS_I(dir)->root;
4250         struct btrfs_trans_handle *trans;
4251         struct inode *inode = d_inode(dentry);
4252         int ret;
4253
4254         trans = __unlink_start_trans(dir);
4255         if (IS_ERR(trans))
4256                 return PTR_ERR(trans);
4257
4258         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4259                         0);
4260
4261         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4262                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4263                         dentry->d_name.len);
4264         if (ret)
4265                 goto out;
4266
4267         if (inode->i_nlink == 0) {
4268                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4269                 if (ret)
4270                         goto out;
4271         }
4272
4273 out:
4274         btrfs_end_transaction(trans);
4275         btrfs_btree_balance_dirty(root->fs_info);
4276         return ret;
4277 }
4278
4279 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4280                         struct btrfs_root *root,
4281                         struct inode *dir, u64 objectid,
4282                         const char *name, int name_len)
4283 {
4284         struct btrfs_fs_info *fs_info = root->fs_info;
4285         struct btrfs_path *path;
4286         struct extent_buffer *leaf;
4287         struct btrfs_dir_item *di;
4288         struct btrfs_key key;
4289         u64 index;
4290         int ret;
4291         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4292
4293         path = btrfs_alloc_path();
4294         if (!path)
4295                 return -ENOMEM;
4296
4297         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4298                                    name, name_len, -1);
4299         if (IS_ERR_OR_NULL(di)) {
4300                 if (!di)
4301                         ret = -ENOENT;
4302                 else
4303                         ret = PTR_ERR(di);
4304                 goto out;
4305         }
4306
4307         leaf = path->nodes[0];
4308         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4309         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4310         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4311         if (ret) {
4312                 btrfs_abort_transaction(trans, ret);
4313                 goto out;
4314         }
4315         btrfs_release_path(path);
4316
4317         ret = btrfs_del_root_ref(trans, fs_info, objectid,
4318                                  root->root_key.objectid, dir_ino,
4319                                  &index, name, name_len);
4320         if (ret < 0) {
4321                 if (ret != -ENOENT) {
4322                         btrfs_abort_transaction(trans, ret);
4323                         goto out;
4324                 }
4325                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4326                                                  name, name_len);
4327                 if (IS_ERR_OR_NULL(di)) {
4328                         if (!di)
4329                                 ret = -ENOENT;
4330                         else
4331                                 ret = PTR_ERR(di);
4332                         btrfs_abort_transaction(trans, ret);
4333                         goto out;
4334                 }
4335
4336                 leaf = path->nodes[0];
4337                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4338                 btrfs_release_path(path);
4339                 index = key.offset;
4340         }
4341         btrfs_release_path(path);
4342
4343         ret = btrfs_delete_delayed_dir_index(trans, fs_info, BTRFS_I(dir), index);
4344         if (ret) {
4345                 btrfs_abort_transaction(trans, ret);
4346                 goto out;
4347         }
4348
4349         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4350         inode_inc_iversion(dir);
4351         dir->i_mtime = dir->i_ctime = current_time(dir);
4352         ret = btrfs_update_inode_fallback(trans, root, dir);
4353         if (ret)
4354                 btrfs_abort_transaction(trans, ret);
4355 out:
4356         btrfs_free_path(path);
4357         return ret;
4358 }
4359
4360 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4361 {
4362         struct inode *inode = d_inode(dentry);
4363         int err = 0;
4364         struct btrfs_root *root = BTRFS_I(dir)->root;
4365         struct btrfs_trans_handle *trans;
4366         u64 last_unlink_trans;
4367
4368         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4369                 return -ENOTEMPTY;
4370         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4371                 return -EPERM;
4372
4373         trans = __unlink_start_trans(dir);
4374         if (IS_ERR(trans))
4375                 return PTR_ERR(trans);
4376
4377         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4378                 err = btrfs_unlink_subvol(trans, root, dir,
4379                                           BTRFS_I(inode)->location.objectid,
4380                                           dentry->d_name.name,
4381                                           dentry->d_name.len);
4382                 goto out;
4383         }
4384
4385         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4386         if (err)
4387                 goto out;
4388
4389         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4390
4391         /* now the directory is empty */
4392         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4393                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4394                         dentry->d_name.len);
4395         if (!err) {
4396                 btrfs_i_size_write(BTRFS_I(inode), 0);
4397                 /*
4398                  * Propagate the last_unlink_trans value of the deleted dir to
4399                  * its parent directory. This is to prevent an unrecoverable
4400                  * log tree in the case we do something like this:
4401                  * 1) create dir foo
4402                  * 2) create snapshot under dir foo
4403                  * 3) delete the snapshot
4404                  * 4) rmdir foo
4405                  * 5) mkdir foo
4406                  * 6) fsync foo or some file inside foo
4407                  */
4408                 if (last_unlink_trans >= trans->transid)
4409                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4410         }
4411 out:
4412         btrfs_end_transaction(trans);
4413         btrfs_btree_balance_dirty(root->fs_info);
4414
4415         return err;
4416 }
4417
4418 static int truncate_space_check(struct btrfs_trans_handle *trans,
4419                                 struct btrfs_root *root,
4420                                 u64 bytes_deleted)
4421 {
4422         struct btrfs_fs_info *fs_info = root->fs_info;
4423         int ret;
4424
4425         /*
4426          * This is only used to apply pressure to the enospc system, we don't
4427          * intend to use this reservation at all.
4428          */
4429         bytes_deleted = btrfs_csum_bytes_to_leaves(fs_info, bytes_deleted);
4430         bytes_deleted *= fs_info->nodesize;
4431         ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
4432                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4433         if (!ret) {
4434                 trace_btrfs_space_reservation(fs_info, "transaction",
4435                                               trans->transid,
4436                                               bytes_deleted, 1);
4437                 trans->bytes_reserved += bytes_deleted;
4438         }
4439         return ret;
4440
4441 }
4442
4443 static int truncate_inline_extent(struct inode *inode,
4444                                   struct btrfs_path *path,
4445                                   struct btrfs_key *found_key,
4446                                   const u64 item_end,
4447                                   const u64 new_size)
4448 {
4449         struct extent_buffer *leaf = path->nodes[0];
4450         int slot = path->slots[0];
4451         struct btrfs_file_extent_item *fi;
4452         u32 size = (u32)(new_size - found_key->offset);
4453         struct btrfs_root *root = BTRFS_I(inode)->root;
4454
4455         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4456
4457         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4458                 loff_t offset = new_size;
4459                 loff_t page_end = ALIGN(offset, PAGE_SIZE);
4460
4461                 /*
4462                  * Zero out the remaining of the last page of our inline extent,
4463                  * instead of directly truncating our inline extent here - that
4464                  * would be much more complex (decompressing all the data, then
4465                  * compressing the truncated data, which might be bigger than
4466                  * the size of the inline extent, resize the extent, etc).
4467                  * We release the path because to get the page we might need to
4468                  * read the extent item from disk (data not in the page cache).
4469                  */
4470                 btrfs_release_path(path);
4471                 return btrfs_truncate_block(inode, offset, page_end - offset,
4472                                         0);
4473         }
4474
4475         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4476         size = btrfs_file_extent_calc_inline_size(size);
4477         btrfs_truncate_item(root->fs_info, path, size, 1);
4478
4479         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4480                 inode_sub_bytes(inode, item_end + 1 - new_size);
4481
4482         return 0;
4483 }
4484
4485 /*
4486  * this can truncate away extent items, csum items and directory items.
4487  * It starts at a high offset and removes keys until it can't find
4488  * any higher than new_size
4489  *
4490  * csum items that cross the new i_size are truncated to the new size
4491  * as well.
4492  *
4493  * min_type is the minimum key type to truncate down to.  If set to 0, this
4494  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4495  */
4496 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4497                                struct btrfs_root *root,
4498                                struct inode *inode,
4499                                u64 new_size, u32 min_type)
4500 {
4501         struct btrfs_fs_info *fs_info = root->fs_info;
4502         struct btrfs_path *path;
4503         struct extent_buffer *leaf;
4504         struct btrfs_file_extent_item *fi;
4505         struct btrfs_key key;
4506         struct btrfs_key found_key;
4507         u64 extent_start = 0;
4508         u64 extent_num_bytes = 0;
4509         u64 extent_offset = 0;
4510         u64 item_end = 0;
4511         u64 last_size = new_size;
4512         u32 found_type = (u8)-1;
4513         int found_extent;
4514         int del_item;
4515         int pending_del_nr = 0;
4516         int pending_del_slot = 0;
4517         int extent_type = -1;
4518         int ret;
4519         int err = 0;
4520         u64 ino = btrfs_ino(BTRFS_I(inode));
4521         u64 bytes_deleted = 0;
4522         bool be_nice = 0;
4523         bool should_throttle = 0;
4524         bool should_end = 0;
4525
4526         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4527
4528         /*
4529          * for non-free space inodes and ref cows, we want to back off from
4530          * time to time
4531          */
4532         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4533             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4534                 be_nice = 1;
4535
4536         path = btrfs_alloc_path();
4537         if (!path)
4538                 return -ENOMEM;
4539         path->reada = READA_BACK;
4540
4541         /*
4542          * We want to drop from the next block forward in case this new size is
4543          * not block aligned since we will be keeping the last block of the
4544          * extent just the way it is.
4545          */
4546         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4547             root == fs_info->tree_root)
4548                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4549                                         fs_info->sectorsize),
4550                                         (u64)-1, 0);
4551
4552         /*
4553          * This function is also used to drop the items in the log tree before
4554          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4555          * it is used to drop the loged items. So we shouldn't kill the delayed
4556          * items.
4557          */
4558         if (min_type == 0 && root == BTRFS_I(inode)->root)
4559                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4560
4561         key.objectid = ino;
4562         key.offset = (u64)-1;
4563         key.type = (u8)-1;
4564
4565 search_again:
4566         /*
4567          * with a 16K leaf size and 128MB extents, you can actually queue
4568          * up a huge file in a single leaf.  Most of the time that
4569          * bytes_deleted is > 0, it will be huge by the time we get here
4570          */
4571         if (be_nice && bytes_deleted > SZ_32M) {
4572                 if (btrfs_should_end_transaction(trans)) {
4573                         err = -EAGAIN;
4574                         goto error;
4575                 }
4576         }
4577
4578
4579         path->leave_spinning = 1;
4580         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4581         if (ret < 0) {
4582                 err = ret;
4583                 goto out;
4584         }
4585
4586         if (ret > 0) {
4587                 /* there are no items in the tree for us to truncate, we're
4588                  * done
4589                  */
4590                 if (path->slots[0] == 0)
4591                         goto out;
4592                 path->slots[0]--;
4593         }
4594
4595         while (1) {
4596                 fi = NULL;
4597                 leaf = path->nodes[0];
4598                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4599                 found_type = found_key.type;
4600
4601                 if (found_key.objectid != ino)
4602                         break;
4603
4604                 if (found_type < min_type)
4605                         break;
4606
4607                 item_end = found_key.offset;
4608                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4609                         fi = btrfs_item_ptr(leaf, path->slots[0],
4610                                             struct btrfs_file_extent_item);
4611                         extent_type = btrfs_file_extent_type(leaf, fi);
4612                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4613                                 item_end +=
4614                                     btrfs_file_extent_num_bytes(leaf, fi);
4615
4616                                 trace_btrfs_truncate_show_fi_regular(
4617                                         BTRFS_I(inode), leaf, fi,
4618                                         found_key.offset);
4619                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4620                                 item_end += btrfs_file_extent_inline_len(leaf,
4621                                                          path->slots[0], fi);
4622
4623                                 trace_btrfs_truncate_show_fi_inline(
4624                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4625                                         found_key.offset);
4626                         }
4627                         item_end--;
4628                 }
4629                 if (found_type > min_type) {
4630                         del_item = 1;
4631                 } else {
4632                         if (item_end < new_size)
4633                                 break;
4634                         if (found_key.offset >= new_size)
4635                                 del_item = 1;
4636                         else
4637                                 del_item = 0;
4638                 }
4639                 found_extent = 0;
4640                 /* FIXME, shrink the extent if the ref count is only 1 */
4641                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4642                         goto delete;
4643
4644                 if (del_item)
4645                         last_size = found_key.offset;
4646                 else
4647                         last_size = new_size;
4648
4649                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4650                         u64 num_dec;
4651                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4652                         if (!del_item) {
4653                                 u64 orig_num_bytes =
4654                                         btrfs_file_extent_num_bytes(leaf, fi);
4655                                 extent_num_bytes = ALIGN(new_size -
4656                                                 found_key.offset,
4657                                                 fs_info->sectorsize);
4658                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4659                                                          extent_num_bytes);
4660                                 num_dec = (orig_num_bytes -
4661                                            extent_num_bytes);
4662                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4663                                              &root->state) &&
4664                                     extent_start != 0)
4665                                         inode_sub_bytes(inode, num_dec);
4666                                 btrfs_mark_buffer_dirty(leaf);
4667                         } else {
4668                                 extent_num_bytes =
4669                                         btrfs_file_extent_disk_num_bytes(leaf,
4670                                                                          fi);
4671                                 extent_offset = found_key.offset -
4672                                         btrfs_file_extent_offset(leaf, fi);
4673
4674                                 /* FIXME blocksize != 4096 */
4675                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4676                                 if (extent_start != 0) {
4677                                         found_extent = 1;
4678                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4679                                                      &root->state))
4680                                                 inode_sub_bytes(inode, num_dec);
4681                                 }
4682                         }
4683                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4684                         /*
4685                          * we can't truncate inline items that have had
4686                          * special encodings
4687                          */
4688                         if (!del_item &&
4689                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4690                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4691
4692                                 /*
4693                                  * Need to release path in order to truncate a
4694                                  * compressed extent. So delete any accumulated
4695                                  * extent items so far.
4696                                  */
4697                                 if (btrfs_file_extent_compression(leaf, fi) !=
4698                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4699                                         err = btrfs_del_items(trans, root, path,
4700                                                               pending_del_slot,
4701                                                               pending_del_nr);
4702                                         if (err) {
4703                                                 btrfs_abort_transaction(trans,
4704                                                                         err);
4705                                                 goto error;
4706                                         }
4707                                         pending_del_nr = 0;
4708                                 }
4709
4710                                 err = truncate_inline_extent(inode, path,
4711                                                              &found_key,
4712                                                              item_end,
4713                                                              new_size);
4714                                 if (err) {
4715                                         btrfs_abort_transaction(trans, err);
4716                                         goto error;
4717                                 }
4718                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4719                                             &root->state)) {
4720                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4721                         }
4722                 }
4723 delete:
4724                 if (del_item) {
4725                         if (!pending_del_nr) {
4726                                 /* no pending yet, add ourselves */
4727                                 pending_del_slot = path->slots[0];
4728                                 pending_del_nr = 1;
4729                         } else if (pending_del_nr &&
4730                                    path->slots[0] + 1 == pending_del_slot) {
4731                                 /* hop on the pending chunk */
4732                                 pending_del_nr++;
4733                                 pending_del_slot = path->slots[0];
4734                         } else {
4735                                 BUG();
4736                         }
4737                 } else {
4738                         break;
4739                 }
4740                 should_throttle = 0;
4741
4742                 if (found_extent &&
4743                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4744                      root == fs_info->tree_root)) {
4745                         btrfs_set_path_blocking(path);
4746                         bytes_deleted += extent_num_bytes;
4747                         ret = btrfs_free_extent(trans, fs_info, extent_start,
4748                                                 extent_num_bytes, 0,
4749                                                 btrfs_header_owner(leaf),
4750                                                 ino, extent_offset);
4751                         if (ret) {
4752                                 btrfs_abort_transaction(trans, ret);
4753                                 break;
4754                         }
4755                         if (btrfs_should_throttle_delayed_refs(trans, fs_info))
4756                                 btrfs_async_run_delayed_refs(fs_info,
4757                                         trans->delayed_ref_updates * 2,
4758                                         trans->transid, 0);
4759                         if (be_nice) {
4760                                 if (truncate_space_check(trans, root,
4761                                                          extent_num_bytes)) {
4762                                         should_end = 1;
4763                                 }
4764                                 if (btrfs_should_throttle_delayed_refs(trans,
4765                                                                        fs_info))
4766                                         should_throttle = 1;
4767                         }
4768                 }
4769
4770                 if (found_type == BTRFS_INODE_ITEM_KEY)
4771                         break;
4772
4773                 if (path->slots[0] == 0 ||
4774                     path->slots[0] != pending_del_slot ||
4775                     should_throttle || should_end) {
4776                         if (pending_del_nr) {
4777                                 ret = btrfs_del_items(trans, root, path,
4778                                                 pending_del_slot,
4779                                                 pending_del_nr);
4780                                 if (ret) {
4781                                         btrfs_abort_transaction(trans, ret);
4782                                         goto error;
4783                                 }
4784                                 pending_del_nr = 0;
4785                         }
4786                         btrfs_release_path(path);
4787                         if (should_throttle) {
4788                                 unsigned long updates = trans->delayed_ref_updates;
4789                                 if (updates) {
4790                                         trans->delayed_ref_updates = 0;
4791                                         ret = btrfs_run_delayed_refs(trans,
4792                                                                    fs_info,
4793                                                                    updates * 2);
4794                                         if (ret && !err)
4795                                                 err = ret;
4796                                 }
4797                         }
4798                         /*
4799                          * if we failed to refill our space rsv, bail out
4800                          * and let the transaction restart
4801                          */
4802                         if (should_end) {
4803                                 err = -EAGAIN;
4804                                 goto error;
4805                         }
4806                         goto search_again;
4807                 } else {
4808                         path->slots[0]--;
4809                 }
4810         }
4811 out:
4812         if (pending_del_nr) {
4813                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4814                                       pending_del_nr);
4815                 if (ret)
4816                         btrfs_abort_transaction(trans, ret);
4817         }
4818 error:
4819         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4820                 ASSERT(last_size >= new_size);
4821                 if (!err && last_size > new_size)
4822                         last_size = new_size;
4823                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4824         }
4825
4826         btrfs_free_path(path);
4827
4828         if (be_nice && bytes_deleted > SZ_32M) {
4829                 unsigned long updates = trans->delayed_ref_updates;
4830                 if (updates) {
4831                         trans->delayed_ref_updates = 0;
4832                         ret = btrfs_run_delayed_refs(trans, fs_info,
4833                                                      updates * 2);
4834                         if (ret && !err)
4835                                 err = ret;
4836                 }
4837         }
4838         return err;
4839 }
4840
4841 /*
4842  * btrfs_truncate_block - read, zero a chunk and write a block
4843  * @inode - inode that we're zeroing
4844  * @from - the offset to start zeroing
4845  * @len - the length to zero, 0 to zero the entire range respective to the
4846  *      offset
4847  * @front - zero up to the offset instead of from the offset on
4848  *
4849  * This will find the block for the "from" offset and cow the block and zero the
4850  * part we want to zero.  This is used with truncate and hole punching.
4851  */
4852 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4853                         int front)
4854 {
4855         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4856         struct address_space *mapping = inode->i_mapping;
4857         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4858         struct btrfs_ordered_extent *ordered;
4859         struct extent_state *cached_state = NULL;
4860         struct extent_changeset *data_reserved = NULL;
4861         char *kaddr;
4862         u32 blocksize = fs_info->sectorsize;
4863         pgoff_t index = from >> PAGE_SHIFT;
4864         unsigned offset = from & (blocksize - 1);
4865         struct page *page;
4866         gfp_t mask = btrfs_alloc_write_mask(mapping);
4867         int ret = 0;
4868         u64 block_start;
4869         u64 block_end;
4870
4871         if ((offset & (blocksize - 1)) == 0 &&
4872             (!len || ((len & (blocksize - 1)) == 0)))
4873                 goto out;
4874
4875         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4876                         round_down(from, blocksize), blocksize);
4877         if (ret)
4878                 goto out;
4879
4880 again:
4881         page = find_or_create_page(mapping, index, mask);
4882         if (!page) {
4883                 btrfs_delalloc_release_space(inode, data_reserved,
4884                                 round_down(from, blocksize),
4885                                 blocksize);
4886                 ret = -ENOMEM;
4887                 goto out;
4888         }
4889
4890         block_start = round_down(from, blocksize);
4891         block_end = block_start + blocksize - 1;
4892
4893         if (!PageUptodate(page)) {
4894                 ret = btrfs_readpage(NULL, page);
4895                 lock_page(page);
4896                 if (page->mapping != mapping) {
4897                         unlock_page(page);
4898                         put_page(page);
4899                         goto again;
4900                 }
4901                 if (!PageUptodate(page)) {
4902                         ret = -EIO;
4903                         goto out_unlock;
4904                 }
4905         }
4906         wait_on_page_writeback(page);
4907
4908         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4909         set_page_extent_mapped(page);
4910
4911         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4912         if (ordered) {
4913                 unlock_extent_cached(io_tree, block_start, block_end,
4914                                      &cached_state, GFP_NOFS);
4915                 unlock_page(page);
4916                 put_page(page);
4917                 btrfs_start_ordered_extent(inode, ordered, 1);
4918                 btrfs_put_ordered_extent(ordered);
4919                 goto again;
4920         }
4921
4922         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4923                           EXTENT_DIRTY | EXTENT_DELALLOC |
4924                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4925                           0, 0, &cached_state, GFP_NOFS);
4926
4927         ret = btrfs_set_extent_delalloc(inode, block_start, block_end,
4928                                         &cached_state, 0);
4929         if (ret) {
4930                 unlock_extent_cached(io_tree, block_start, block_end,
4931                                      &cached_state, GFP_NOFS);
4932                 goto out_unlock;
4933         }
4934
4935         if (offset != blocksize) {
4936                 if (!len)
4937                         len = blocksize - offset;
4938                 kaddr = kmap(page);
4939                 if (front)
4940                         memset(kaddr + (block_start - page_offset(page)),
4941                                 0, offset);
4942                 else
4943                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4944                                 0, len);
4945                 flush_dcache_page(page);
4946                 kunmap(page);
4947         }
4948         ClearPageChecked(page);
4949         set_page_dirty(page);
4950         unlock_extent_cached(io_tree, block_start, block_end, &cached_state,
4951                              GFP_NOFS);
4952
4953 out_unlock:
4954         if (ret)
4955                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
4956                                              blocksize);
4957         unlock_page(page);
4958         put_page(page);
4959 out:
4960         extent_changeset_free(data_reserved);
4961         return ret;
4962 }
4963
4964 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4965                              u64 offset, u64 len)
4966 {
4967         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4968         struct btrfs_trans_handle *trans;
4969         int ret;
4970
4971         /*
4972          * Still need to make sure the inode looks like it's been updated so
4973          * that any holes get logged if we fsync.
4974          */
4975         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4976                 BTRFS_I(inode)->last_trans = fs_info->generation;
4977                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4978                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4979                 return 0;
4980         }
4981
4982         /*
4983          * 1 - for the one we're dropping
4984          * 1 - for the one we're adding
4985          * 1 - for updating the inode.
4986          */
4987         trans = btrfs_start_transaction(root, 3);
4988         if (IS_ERR(trans))
4989                 return PTR_ERR(trans);
4990
4991         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4992         if (ret) {
4993                 btrfs_abort_transaction(trans, ret);
4994                 btrfs_end_transaction(trans);
4995                 return ret;
4996         }
4997
4998         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4999                         offset, 0, 0, len, 0, len, 0, 0, 0);
5000         if (ret)
5001                 btrfs_abort_transaction(trans, ret);
5002         else
5003                 btrfs_update_inode(trans, root, inode);
5004         btrfs_end_transaction(trans);
5005         return ret;
5006 }
5007
5008 /*
5009  * This function puts in dummy file extents for the area we're creating a hole
5010  * for.  So if we are truncating this file to a larger size we need to insert
5011  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
5012  * the range between oldsize and size
5013  */
5014 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
5015 {
5016         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5017         struct btrfs_root *root = BTRFS_I(inode)->root;
5018         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5019         struct extent_map *em = NULL;
5020         struct extent_state *cached_state = NULL;
5021         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
5022         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
5023         u64 block_end = ALIGN(size, fs_info->sectorsize);
5024         u64 last_byte;
5025         u64 cur_offset;
5026         u64 hole_size;
5027         int err = 0;
5028
5029         /*
5030          * If our size started in the middle of a block we need to zero out the
5031          * rest of the block before we expand the i_size, otherwise we could
5032          * expose stale data.
5033          */
5034         err = btrfs_truncate_block(inode, oldsize, 0, 0);
5035         if (err)
5036                 return err;
5037
5038         if (size <= hole_start)
5039                 return 0;
5040
5041         while (1) {
5042                 struct btrfs_ordered_extent *ordered;
5043
5044                 lock_extent_bits(io_tree, hole_start, block_end - 1,
5045                                  &cached_state);
5046                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), hole_start,
5047                                                      block_end - hole_start);
5048                 if (!ordered)
5049                         break;
5050                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
5051                                      &cached_state, GFP_NOFS);
5052                 btrfs_start_ordered_extent(inode, ordered, 1);
5053                 btrfs_put_ordered_extent(ordered);
5054         }
5055
5056         cur_offset = hole_start;
5057         while (1) {
5058                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
5059                                 block_end - cur_offset, 0);
5060                 if (IS_ERR(em)) {
5061                         err = PTR_ERR(em);
5062                         em = NULL;
5063                         break;
5064                 }
5065                 last_byte = min(extent_map_end(em), block_end);
5066                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
5067                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
5068                         struct extent_map *hole_em;
5069                         hole_size = last_byte - cur_offset;
5070
5071                         err = maybe_insert_hole(root, inode, cur_offset,
5072                                                 hole_size);
5073                         if (err)
5074                                 break;
5075                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
5076                                                 cur_offset + hole_size - 1, 0);
5077                         hole_em = alloc_extent_map();
5078                         if (!hole_em) {
5079                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5080                                         &BTRFS_I(inode)->runtime_flags);
5081                                 goto next;
5082                         }
5083                         hole_em->start = cur_offset;
5084                         hole_em->len = hole_size;
5085                         hole_em->orig_start = cur_offset;
5086
5087                         hole_em->block_start = EXTENT_MAP_HOLE;
5088                         hole_em->block_len = 0;
5089                         hole_em->orig_block_len = 0;
5090                         hole_em->ram_bytes = hole_size;
5091                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
5092                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
5093                         hole_em->generation = fs_info->generation;
5094
5095                         while (1) {
5096                                 write_lock(&em_tree->lock);
5097                                 err = add_extent_mapping(em_tree, hole_em, 1);
5098                                 write_unlock(&em_tree->lock);
5099                                 if (err != -EEXIST)
5100                                         break;
5101                                 btrfs_drop_extent_cache(BTRFS_I(inode),
5102                                                         cur_offset,
5103                                                         cur_offset +
5104                                                         hole_size - 1, 0);
5105                         }
5106                         free_extent_map(hole_em);
5107                 }
5108 next:
5109                 free_extent_map(em);
5110                 em = NULL;
5111                 cur_offset = last_byte;
5112                 if (cur_offset >= block_end)
5113                         break;
5114         }
5115         free_extent_map(em);
5116         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
5117                              GFP_NOFS);
5118         return err;
5119 }
5120
5121 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5122 {
5123         struct btrfs_root *root = BTRFS_I(inode)->root;
5124         struct btrfs_trans_handle *trans;
5125         loff_t oldsize = i_size_read(inode);
5126         loff_t newsize = attr->ia_size;
5127         int mask = attr->ia_valid;
5128         int ret;
5129
5130         /*
5131          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5132          * special case where we need to update the times despite not having
5133          * these flags set.  For all other operations the VFS set these flags
5134          * explicitly if it wants a timestamp update.
5135          */
5136         if (newsize != oldsize) {
5137                 inode_inc_iversion(inode);
5138                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5139                         inode->i_ctime = inode->i_mtime =
5140                                 current_time(inode);
5141         }
5142
5143         if (newsize > oldsize) {
5144                 /*
5145                  * Don't do an expanding truncate while snapshotting is ongoing.
5146                  * This is to ensure the snapshot captures a fully consistent
5147                  * state of this file - if the snapshot captures this expanding
5148                  * truncation, it must capture all writes that happened before
5149                  * this truncation.
5150                  */
5151                 btrfs_wait_for_snapshot_creation(root);
5152                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5153                 if (ret) {
5154                         btrfs_end_write_no_snapshotting(root);
5155                         return ret;
5156                 }
5157
5158                 trans = btrfs_start_transaction(root, 1);
5159                 if (IS_ERR(trans)) {
5160                         btrfs_end_write_no_snapshotting(root);
5161                         return PTR_ERR(trans);
5162                 }
5163
5164                 i_size_write(inode, newsize);
5165                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5166                 pagecache_isize_extended(inode, oldsize, newsize);
5167                 ret = btrfs_update_inode(trans, root, inode);
5168                 btrfs_end_write_no_snapshotting(root);
5169                 btrfs_end_transaction(trans);
5170         } else {
5171
5172                 /*
5173                  * We're truncating a file that used to have good data down to
5174                  * zero. Make sure it gets into the ordered flush list so that
5175                  * any new writes get down to disk quickly.
5176                  */
5177                 if (newsize == 0)
5178                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5179                                 &BTRFS_I(inode)->runtime_flags);
5180
5181                 /*
5182                  * 1 for the orphan item we're going to add
5183                  * 1 for the orphan item deletion.
5184                  */
5185                 trans = btrfs_start_transaction(root, 2);
5186                 if (IS_ERR(trans))
5187                         return PTR_ERR(trans);
5188
5189                 /*
5190                  * We need to do this in case we fail at _any_ point during the
5191                  * actual truncate.  Once we do the truncate_setsize we could
5192                  * invalidate pages which forces any outstanding ordered io to
5193                  * be instantly completed which will give us extents that need
5194                  * to be truncated.  If we fail to get an orphan inode down we
5195                  * could have left over extents that were never meant to live,
5196                  * so we need to guarantee from this point on that everything
5197                  * will be consistent.
5198                  */
5199                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
5200                 btrfs_end_transaction(trans);
5201                 if (ret)
5202                         return ret;
5203
5204                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
5205                 truncate_setsize(inode, newsize);
5206
5207                 /* Disable nonlocked read DIO to avoid the end less truncate */
5208                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5209                 inode_dio_wait(inode);
5210                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5211
5212                 ret = btrfs_truncate(inode);
5213                 if (ret && inode->i_nlink) {
5214                         int err;
5215
5216                         /* To get a stable disk_i_size */
5217                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5218                         if (err) {
5219                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5220                                 return err;
5221                         }
5222
5223                         /*
5224                          * failed to truncate, disk_i_size is only adjusted down
5225                          * as we remove extents, so it should represent the true
5226                          * size of the inode, so reset the in memory size and
5227                          * delete our orphan entry.
5228                          */
5229                         trans = btrfs_join_transaction(root);
5230                         if (IS_ERR(trans)) {
5231                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5232                                 return ret;
5233                         }
5234                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5235                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
5236                         if (err)
5237                                 btrfs_abort_transaction(trans, err);
5238                         btrfs_end_transaction(trans);
5239                 }
5240         }
5241
5242         return ret;
5243 }
5244
5245 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5246 {
5247         struct inode *inode = d_inode(dentry);
5248         struct btrfs_root *root = BTRFS_I(inode)->root;
5249         int err;
5250
5251         if (btrfs_root_readonly(root))
5252                 return -EROFS;
5253
5254         err = setattr_prepare(dentry, attr);
5255         if (err)
5256                 return err;
5257
5258         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5259                 err = btrfs_setsize(inode, attr);
5260                 if (err)
5261                         return err;
5262         }
5263
5264         if (attr->ia_valid) {
5265                 setattr_copy(inode, attr);
5266                 inode_inc_iversion(inode);
5267                 err = btrfs_dirty_inode(inode);
5268
5269                 if (!err && attr->ia_valid & ATTR_MODE)
5270                         err = posix_acl_chmod(inode, inode->i_mode);
5271         }
5272
5273         return err;
5274 }
5275
5276 /*
5277  * While truncating the inode pages during eviction, we get the VFS calling
5278  * btrfs_invalidatepage() against each page of the inode. This is slow because
5279  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5280  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5281  * extent_state structures over and over, wasting lots of time.
5282  *
5283  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5284  * those expensive operations on a per page basis and do only the ordered io
5285  * finishing, while we release here the extent_map and extent_state structures,
5286  * without the excessive merging and splitting.
5287  */
5288 static void evict_inode_truncate_pages(struct inode *inode)
5289 {
5290         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5291         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5292         struct rb_node *node;
5293
5294         ASSERT(inode->i_state & I_FREEING);
5295         truncate_inode_pages_final(&inode->i_data);
5296
5297         write_lock(&map_tree->lock);
5298         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5299                 struct extent_map *em;
5300
5301                 node = rb_first(&map_tree->map);
5302                 em = rb_entry(node, struct extent_map, rb_node);
5303                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5304                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5305                 remove_extent_mapping(map_tree, em);
5306                 free_extent_map(em);
5307                 if (need_resched()) {
5308                         write_unlock(&map_tree->lock);
5309                         cond_resched();
5310                         write_lock(&map_tree->lock);
5311                 }
5312         }
5313         write_unlock(&map_tree->lock);
5314
5315         /*
5316          * Keep looping until we have no more ranges in the io tree.
5317          * We can have ongoing bios started by readpages (called from readahead)
5318          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5319          * still in progress (unlocked the pages in the bio but did not yet
5320          * unlocked the ranges in the io tree). Therefore this means some
5321          * ranges can still be locked and eviction started because before
5322          * submitting those bios, which are executed by a separate task (work
5323          * queue kthread), inode references (inode->i_count) were not taken
5324          * (which would be dropped in the end io callback of each bio).
5325          * Therefore here we effectively end up waiting for those bios and
5326          * anyone else holding locked ranges without having bumped the inode's
5327          * reference count - if we don't do it, when they access the inode's
5328          * io_tree to unlock a range it may be too late, leading to an
5329          * use-after-free issue.
5330          */
5331         spin_lock(&io_tree->lock);
5332         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5333                 struct extent_state *state;
5334                 struct extent_state *cached_state = NULL;
5335                 u64 start;
5336                 u64 end;
5337                 unsigned state_flags;
5338
5339                 node = rb_first(&io_tree->state);
5340                 state = rb_entry(node, struct extent_state, rb_node);
5341                 start = state->start;
5342                 end = state->end;
5343                 state_flags = state->state;
5344                 spin_unlock(&io_tree->lock);
5345
5346                 lock_extent_bits(io_tree, start, end, &cached_state);
5347
5348                 /*
5349                  * If still has DELALLOC flag, the extent didn't reach disk,
5350                  * and its reserved space won't be freed by delayed_ref.
5351                  * So we need to free its reserved space here.
5352                  * (Refer to comment in btrfs_invalidatepage, case 2)
5353                  *
5354                  * Note, end is the bytenr of last byte, so we need + 1 here.
5355                  */
5356                 if (state_flags & EXTENT_DELALLOC)
5357                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5358
5359                 clear_extent_bit(io_tree, start, end,
5360                                  EXTENT_LOCKED | EXTENT_DIRTY |
5361                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5362                                  EXTENT_DEFRAG, 1, 1,
5363                                  &cached_state, GFP_NOFS);
5364
5365                 cond_resched();
5366                 spin_lock(&io_tree->lock);
5367         }
5368         spin_unlock(&io_tree->lock);
5369 }
5370
5371 void btrfs_evict_inode(struct inode *inode)
5372 {
5373         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5374         struct btrfs_trans_handle *trans;
5375         struct btrfs_root *root = BTRFS_I(inode)->root;
5376         struct btrfs_block_rsv *rsv, *global_rsv;
5377         int steal_from_global = 0;
5378         u64 min_size;
5379         int ret;
5380
5381         trace_btrfs_inode_evict(inode);
5382
5383         if (!root) {
5384                 clear_inode(inode);
5385                 return;
5386         }
5387
5388         min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
5389
5390         evict_inode_truncate_pages(inode);
5391
5392         if (inode->i_nlink &&
5393             ((btrfs_root_refs(&root->root_item) != 0 &&
5394               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5395              btrfs_is_free_space_inode(BTRFS_I(inode))))
5396                 goto no_delete;
5397
5398         if (is_bad_inode(inode)) {
5399                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5400                 goto no_delete;
5401         }
5402         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5403         if (!special_file(inode->i_mode))
5404                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5405
5406         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5407
5408         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
5409                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5410                                  &BTRFS_I(inode)->runtime_flags));
5411                 goto no_delete;
5412         }
5413
5414         if (inode->i_nlink > 0) {
5415                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5416                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5417                 goto no_delete;
5418         }
5419
5420         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5421         if (ret) {
5422                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5423                 goto no_delete;
5424         }
5425
5426         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5427         if (!rsv) {
5428                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5429                 goto no_delete;
5430         }
5431         rsv->size = min_size;
5432         rsv->failfast = 1;
5433         global_rsv = &fs_info->global_block_rsv;
5434
5435         btrfs_i_size_write(BTRFS_I(inode), 0);
5436
5437         /*
5438          * This is a bit simpler than btrfs_truncate since we've already
5439          * reserved our space for our orphan item in the unlink, so we just
5440          * need to reserve some slack space in case we add bytes and update
5441          * inode item when doing the truncate.
5442          */
5443         while (1) {
5444                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5445                                              BTRFS_RESERVE_FLUSH_LIMIT);
5446
5447                 /*
5448                  * Try and steal from the global reserve since we will
5449                  * likely not use this space anyway, we want to try as
5450                  * hard as possible to get this to work.
5451                  */
5452                 if (ret)
5453                         steal_from_global++;
5454                 else
5455                         steal_from_global = 0;
5456                 ret = 0;
5457
5458                 /*
5459                  * steal_from_global == 0: we reserved stuff, hooray!
5460                  * steal_from_global == 1: we didn't reserve stuff, boo!
5461                  * steal_from_global == 2: we've committed, still not a lot of
5462                  * room but maybe we'll have room in the global reserve this
5463                  * time.
5464                  * steal_from_global == 3: abandon all hope!
5465                  */
5466                 if (steal_from_global > 2) {
5467                         btrfs_warn(fs_info,
5468                                    "Could not get space for a delete, will truncate on mount %d",
5469                                    ret);
5470                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5471                         btrfs_free_block_rsv(fs_info, rsv);
5472                         goto no_delete;
5473                 }
5474
5475                 trans = btrfs_join_transaction(root);
5476                 if (IS_ERR(trans)) {
5477                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5478                         btrfs_free_block_rsv(fs_info, rsv);
5479                         goto no_delete;
5480                 }
5481
5482                 /*
5483                  * We can't just steal from the global reserve, we need to make
5484                  * sure there is room to do it, if not we need to commit and try
5485                  * again.
5486                  */
5487                 if (steal_from_global) {
5488                         if (!btrfs_check_space_for_delayed_refs(trans, fs_info))
5489                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5490                                                               min_size, 0);
5491                         else
5492                                 ret = -ENOSPC;
5493                 }
5494
5495                 /*
5496                  * Couldn't steal from the global reserve, we have too much
5497                  * pending stuff built up, commit the transaction and try it
5498                  * again.
5499                  */
5500                 if (ret) {
5501                         ret = btrfs_commit_transaction(trans);
5502                         if (ret) {
5503                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5504                                 btrfs_free_block_rsv(fs_info, rsv);
5505                                 goto no_delete;
5506                         }
5507                         continue;
5508                 } else {
5509                         steal_from_global = 0;
5510                 }
5511
5512                 trans->block_rsv = rsv;
5513
5514                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5515                 if (ret) {
5516                         trans->block_rsv = &fs_info->trans_block_rsv;
5517                         btrfs_end_transaction(trans);
5518                         btrfs_btree_balance_dirty(fs_info);
5519                         if (ret != -ENOSPC && ret != -EAGAIN) {
5520                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5521                                 btrfs_free_block_rsv(fs_info, rsv);
5522                                 goto no_delete;
5523                         }
5524                 } else {
5525                         break;
5526                 }
5527         }
5528
5529         btrfs_free_block_rsv(fs_info, rsv);
5530
5531         /*
5532          * Errors here aren't a big deal, it just means we leave orphan items
5533          * in the tree.  They will be cleaned up on the next mount.
5534          */
5535         trans->block_rsv = root->orphan_block_rsv;
5536         btrfs_orphan_del(trans, BTRFS_I(inode));
5537
5538         trans->block_rsv = &fs_info->trans_block_rsv;
5539         if (!(root == fs_info->tree_root ||
5540               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5541                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5542
5543         btrfs_end_transaction(trans);
5544         btrfs_btree_balance_dirty(fs_info);
5545 no_delete:
5546         btrfs_remove_delayed_node(BTRFS_I(inode));
5547         clear_inode(inode);
5548 }
5549
5550 /*
5551  * this returns the key found in the dir entry in the location pointer.
5552  * If no dir entries were found, location->objectid is 0.
5553  */
5554 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5555                                struct btrfs_key *location)
5556 {
5557         const char *name = dentry->d_name.name;
5558         int namelen = dentry->d_name.len;
5559         struct btrfs_dir_item *di;
5560         struct btrfs_path *path;
5561         struct btrfs_root *root = BTRFS_I(dir)->root;
5562         int ret = 0;
5563
5564         path = btrfs_alloc_path();
5565         if (!path)
5566                 return -ENOMEM;
5567
5568         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5569                         name, namelen, 0);
5570         if (IS_ERR(di))
5571                 ret = PTR_ERR(di);
5572
5573         if (IS_ERR_OR_NULL(di))
5574                 goto out_err;
5575
5576         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5577         if (location->type != BTRFS_INODE_ITEM_KEY &&
5578             location->type != BTRFS_ROOT_ITEM_KEY) {
5579                 btrfs_warn(root->fs_info,
5580 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5581                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5582                            location->objectid, location->type, location->offset);
5583                 goto out_err;
5584         }
5585 out:
5586         btrfs_free_path(path);
5587         return ret;
5588 out_err:
5589         location->objectid = 0;
5590         goto out;
5591 }
5592
5593 /*
5594  * when we hit a tree root in a directory, the btrfs part of the inode
5595  * needs to be changed to reflect the root directory of the tree root.  This
5596  * is kind of like crossing a mount point.
5597  */
5598 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5599                                     struct inode *dir,
5600                                     struct dentry *dentry,
5601                                     struct btrfs_key *location,
5602                                     struct btrfs_root **sub_root)
5603 {
5604         struct btrfs_path *path;
5605         struct btrfs_root *new_root;
5606         struct btrfs_root_ref *ref;
5607         struct extent_buffer *leaf;
5608         struct btrfs_key key;
5609         int ret;
5610         int err = 0;
5611
5612         path = btrfs_alloc_path();
5613         if (!path) {
5614                 err = -ENOMEM;
5615                 goto out;
5616         }
5617
5618         err = -ENOENT;
5619         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5620         key.type = BTRFS_ROOT_REF_KEY;
5621         key.offset = location->objectid;
5622
5623         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5624         if (ret) {
5625                 if (ret < 0)
5626                         err = ret;
5627                 goto out;
5628         }
5629
5630         leaf = path->nodes[0];
5631         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5632         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5633             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5634                 goto out;
5635
5636         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5637                                    (unsigned long)(ref + 1),
5638                                    dentry->d_name.len);
5639         if (ret)
5640                 goto out;
5641
5642         btrfs_release_path(path);
5643
5644         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5645         if (IS_ERR(new_root)) {
5646                 err = PTR_ERR(new_root);
5647                 goto out;
5648         }
5649
5650         *sub_root = new_root;
5651         location->objectid = btrfs_root_dirid(&new_root->root_item);
5652         location->type = BTRFS_INODE_ITEM_KEY;
5653         location->offset = 0;
5654         err = 0;
5655 out:
5656         btrfs_free_path(path);
5657         return err;
5658 }
5659
5660 static void inode_tree_add(struct inode *inode)
5661 {
5662         struct btrfs_root *root = BTRFS_I(inode)->root;
5663         struct btrfs_inode *entry;
5664         struct rb_node **p;
5665         struct rb_node *parent;
5666         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5667         u64 ino = btrfs_ino(BTRFS_I(inode));
5668
5669         if (inode_unhashed(inode))
5670                 return;
5671         parent = NULL;
5672         spin_lock(&root->inode_lock);
5673         p = &root->inode_tree.rb_node;
5674         while (*p) {
5675                 parent = *p;
5676                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5677
5678                 if (ino < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5679                         p = &parent->rb_left;
5680                 else if (ino > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5681                         p = &parent->rb_right;
5682                 else {
5683                         WARN_ON(!(entry->vfs_inode.i_state &
5684                                   (I_WILL_FREE | I_FREEING)));
5685                         rb_replace_node(parent, new, &root->inode_tree);
5686                         RB_CLEAR_NODE(parent);
5687                         spin_unlock(&root->inode_lock);
5688                         return;
5689                 }
5690         }
5691         rb_link_node(new, parent, p);
5692         rb_insert_color(new, &root->inode_tree);
5693         spin_unlock(&root->inode_lock);
5694 }
5695
5696 static void inode_tree_del(struct inode *inode)
5697 {
5698         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5699         struct btrfs_root *root = BTRFS_I(inode)->root;
5700         int empty = 0;
5701
5702         spin_lock(&root->inode_lock);
5703         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5704                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5705                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5706                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5707         }
5708         spin_unlock(&root->inode_lock);
5709
5710         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5711                 synchronize_srcu(&fs_info->subvol_srcu);
5712                 spin_lock(&root->inode_lock);
5713                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5714                 spin_unlock(&root->inode_lock);
5715                 if (empty)
5716                         btrfs_add_dead_root(root);
5717         }
5718 }
5719
5720 void btrfs_invalidate_inodes(struct btrfs_root *root)
5721 {
5722         struct btrfs_fs_info *fs_info = root->fs_info;
5723         struct rb_node *node;
5724         struct rb_node *prev;
5725         struct btrfs_inode *entry;
5726         struct inode *inode;
5727         u64 objectid = 0;
5728
5729         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
5730                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5731
5732         spin_lock(&root->inode_lock);
5733 again:
5734         node = root->inode_tree.rb_node;
5735         prev = NULL;
5736         while (node) {
5737                 prev = node;
5738                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5739
5740                 if (objectid < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5741                         node = node->rb_left;
5742                 else if (objectid > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5743                         node = node->rb_right;
5744                 else
5745                         break;
5746         }
5747         if (!node) {
5748                 while (prev) {
5749                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5750                         if (objectid <= btrfs_ino(BTRFS_I(&entry->vfs_inode))) {
5751                                 node = prev;
5752                                 break;
5753                         }
5754                         prev = rb_next(prev);
5755                 }
5756         }
5757         while (node) {
5758                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5759                 objectid = btrfs_ino(BTRFS_I(&entry->vfs_inode)) + 1;
5760                 inode = igrab(&entry->vfs_inode);
5761                 if (inode) {
5762                         spin_unlock(&root->inode_lock);
5763                         if (atomic_read(&inode->i_count) > 1)
5764                                 d_prune_aliases(inode);
5765                         /*
5766                          * btrfs_drop_inode will have it removed from
5767                          * the inode cache when its usage count
5768                          * hits zero.
5769                          */
5770                         iput(inode);
5771                         cond_resched();
5772                         spin_lock(&root->inode_lock);
5773                         goto again;
5774                 }
5775
5776                 if (cond_resched_lock(&root->inode_lock))
5777                         goto again;
5778
5779                 node = rb_next(node);
5780         }
5781         spin_unlock(&root->inode_lock);
5782 }
5783
5784 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5785 {
5786         struct btrfs_iget_args *args = p;
5787         inode->i_ino = args->location->objectid;
5788         memcpy(&BTRFS_I(inode)->location, args->location,
5789                sizeof(*args->location));
5790         BTRFS_I(inode)->root = args->root;
5791         return 0;
5792 }
5793
5794 static int btrfs_find_actor(struct inode *inode, void *opaque)
5795 {
5796         struct btrfs_iget_args *args = opaque;
5797         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5798                 args->root == BTRFS_I(inode)->root;
5799 }
5800
5801 static struct inode *btrfs_iget_locked(struct super_block *s,
5802                                        struct btrfs_key *location,
5803                                        struct btrfs_root *root)
5804 {
5805         struct inode *inode;
5806         struct btrfs_iget_args args;
5807         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5808
5809         args.location = location;
5810         args.root = root;
5811
5812         inode = iget5_locked(s, hashval, btrfs_find_actor,
5813                              btrfs_init_locked_inode,
5814                              (void *)&args);
5815         return inode;
5816 }
5817
5818 /* Get an inode object given its location and corresponding root.
5819  * Returns in *is_new if the inode was read from disk
5820  */
5821 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5822                          struct btrfs_root *root, int *new)
5823 {
5824         struct inode *inode;
5825
5826         inode = btrfs_iget_locked(s, location, root);
5827         if (!inode)
5828                 return ERR_PTR(-ENOMEM);
5829
5830         if (inode->i_state & I_NEW) {
5831                 int ret;
5832
5833                 ret = btrfs_read_locked_inode(inode);
5834                 if (!is_bad_inode(inode)) {
5835                         inode_tree_add(inode);
5836                         unlock_new_inode(inode);
5837                         if (new)
5838                                 *new = 1;
5839                 } else {
5840                         unlock_new_inode(inode);
5841                         iput(inode);
5842                         ASSERT(ret < 0);
5843                         inode = ERR_PTR(ret < 0 ? ret : -ESTALE);
5844                 }
5845         }
5846
5847         return inode;
5848 }
5849
5850 static struct inode *new_simple_dir(struct super_block *s,
5851                                     struct btrfs_key *key,
5852                                     struct btrfs_root *root)
5853 {
5854         struct inode *inode = new_inode(s);
5855
5856         if (!inode)
5857                 return ERR_PTR(-ENOMEM);
5858
5859         BTRFS_I(inode)->root = root;
5860         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5861         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5862
5863         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5864         inode->i_op = &btrfs_dir_ro_inode_operations;
5865         inode->i_opflags &= ~IOP_XATTR;
5866         inode->i_fop = &simple_dir_operations;
5867         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5868         inode->i_mtime = current_time(inode);
5869         inode->i_atime = inode->i_mtime;
5870         inode->i_ctime = inode->i_mtime;
5871         BTRFS_I(inode)->i_otime = inode->i_mtime;
5872
5873         return inode;
5874 }
5875
5876 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5877 {
5878         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5879         struct inode *inode;
5880         struct btrfs_root *root = BTRFS_I(dir)->root;
5881         struct btrfs_root *sub_root = root;
5882         struct btrfs_key location;
5883         int index;
5884         int ret = 0;
5885
5886         if (dentry->d_name.len > BTRFS_NAME_LEN)
5887                 return ERR_PTR(-ENAMETOOLONG);
5888
5889         ret = btrfs_inode_by_name(dir, dentry, &location);
5890         if (ret < 0)
5891                 return ERR_PTR(ret);
5892
5893         if (location.objectid == 0)
5894                 return ERR_PTR(-ENOENT);
5895
5896         if (location.type == BTRFS_INODE_ITEM_KEY) {
5897                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5898                 return inode;
5899         }
5900
5901         index = srcu_read_lock(&fs_info->subvol_srcu);
5902         ret = fixup_tree_root_location(fs_info, dir, dentry,
5903                                        &location, &sub_root);
5904         if (ret < 0) {
5905                 if (ret != -ENOENT)
5906                         inode = ERR_PTR(ret);
5907                 else
5908                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5909         } else {
5910                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5911         }
5912         srcu_read_unlock(&fs_info->subvol_srcu, index);
5913
5914         if (!IS_ERR(inode) && root != sub_root) {
5915                 down_read(&fs_info->cleanup_work_sem);
5916                 if (!sb_rdonly(inode->i_sb))
5917                         ret = btrfs_orphan_cleanup(sub_root);
5918                 up_read(&fs_info->cleanup_work_sem);
5919                 if (ret) {
5920                         iput(inode);
5921                         inode = ERR_PTR(ret);
5922                 }
5923         }
5924
5925         return inode;
5926 }
5927
5928 static int btrfs_dentry_delete(const struct dentry *dentry)
5929 {
5930         struct btrfs_root *root;
5931         struct inode *inode = d_inode(dentry);
5932
5933         if (!inode && !IS_ROOT(dentry))
5934                 inode = d_inode(dentry->d_parent);
5935
5936         if (inode) {
5937                 root = BTRFS_I(inode)->root;
5938                 if (btrfs_root_refs(&root->root_item) == 0)
5939                         return 1;
5940
5941                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5942                         return 1;
5943         }
5944         return 0;
5945 }
5946
5947 static void btrfs_dentry_release(struct dentry *dentry)
5948 {
5949         kfree(dentry->d_fsdata);
5950 }
5951
5952 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5953                                    unsigned int flags)
5954 {
5955         struct inode *inode;
5956
5957         inode = btrfs_lookup_dentry(dir, dentry);
5958         if (IS_ERR(inode)) {
5959                 if (PTR_ERR(inode) == -ENOENT)
5960                         inode = NULL;
5961                 else
5962                         return ERR_CAST(inode);
5963         }
5964
5965         return d_splice_alias(inode, dentry);
5966 }
5967
5968 unsigned char btrfs_filetype_table[] = {
5969         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5970 };
5971
5972 /*
5973  * All this infrastructure exists because dir_emit can fault, and we are holding
5974  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5975  * our information into that, and then dir_emit from the buffer.  This is
5976  * similar to what NFS does, only we don't keep the buffer around in pagecache
5977  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5978  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5979  * tree lock.
5980  */
5981 static int btrfs_opendir(struct inode *inode, struct file *file)
5982 {
5983         struct btrfs_file_private *private;
5984
5985         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5986         if (!private)
5987                 return -ENOMEM;
5988         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5989         if (!private->filldir_buf) {
5990                 kfree(private);
5991                 return -ENOMEM;
5992         }
5993         file->private_data = private;
5994         return 0;
5995 }
5996
5997 struct dir_entry {
5998         u64 ino;
5999         u64 offset;
6000         unsigned type;
6001         int name_len;
6002 };
6003
6004 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
6005 {
6006         while (entries--) {
6007                 struct dir_entry *entry = addr;
6008                 char *name = (char *)(entry + 1);
6009
6010                 ctx->pos = get_unaligned(&entry->offset);
6011                 if (!dir_emit(ctx, name, get_unaligned(&entry->name_len),
6012                                          get_unaligned(&entry->ino),
6013                                          get_unaligned(&entry->type)))
6014                         return 1;
6015                 addr += sizeof(struct dir_entry) +
6016                         get_unaligned(&entry->name_len);
6017                 ctx->pos++;
6018         }
6019         return 0;
6020 }
6021
6022 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
6023 {
6024         struct inode *inode = file_inode(file);
6025         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6026         struct btrfs_root *root = BTRFS_I(inode)->root;
6027         struct btrfs_file_private *private = file->private_data;
6028         struct btrfs_dir_item *di;
6029         struct btrfs_key key;
6030         struct btrfs_key found_key;
6031         struct btrfs_path *path;
6032         void *addr;
6033         struct list_head ins_list;
6034         struct list_head del_list;
6035         int ret;
6036         struct extent_buffer *leaf;
6037         int slot;
6038         char *name_ptr;
6039         int name_len;
6040         int entries = 0;
6041         int total_len = 0;
6042         bool put = false;
6043         struct btrfs_key location;
6044
6045         if (!dir_emit_dots(file, ctx))
6046                 return 0;
6047
6048         path = btrfs_alloc_path();
6049         if (!path)
6050                 return -ENOMEM;
6051
6052         addr = private->filldir_buf;
6053         path->reada = READA_FORWARD;
6054
6055         INIT_LIST_HEAD(&ins_list);
6056         INIT_LIST_HEAD(&del_list);
6057         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
6058
6059 again:
6060         key.type = BTRFS_DIR_INDEX_KEY;
6061         key.offset = ctx->pos;
6062         key.objectid = btrfs_ino(BTRFS_I(inode));
6063
6064         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6065         if (ret < 0)
6066                 goto err;
6067
6068         while (1) {
6069                 struct dir_entry *entry;
6070
6071                 leaf = path->nodes[0];
6072                 slot = path->slots[0];
6073                 if (slot >= btrfs_header_nritems(leaf)) {
6074                         ret = btrfs_next_leaf(root, path);
6075                         if (ret < 0)
6076                                 goto err;
6077                         else if (ret > 0)
6078                                 break;
6079                         continue;
6080                 }
6081
6082                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
6083
6084                 if (found_key.objectid != key.objectid)
6085                         break;
6086                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
6087                         break;
6088                 if (found_key.offset < ctx->pos)
6089                         goto next;
6090                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
6091                         goto next;
6092                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
6093                 if (verify_dir_item(fs_info, leaf, slot, di))
6094                         goto next;
6095
6096                 name_len = btrfs_dir_name_len(leaf, di);
6097                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
6098                     PAGE_SIZE) {
6099                         btrfs_release_path(path);
6100                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6101                         if (ret)
6102                                 goto nopos;
6103                         addr = private->filldir_buf;
6104                         entries = 0;
6105                         total_len = 0;
6106                         goto again;
6107                 }
6108
6109                 entry = addr;
6110                 put_unaligned(name_len, &entry->name_len);
6111                 name_ptr = (char *)(entry + 1);
6112                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
6113                                    name_len);
6114                 put_unaligned(btrfs_filetype_table[btrfs_dir_type(leaf, di)],
6115                                 &entry->type);
6116                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
6117                 put_unaligned(location.objectid, &entry->ino);
6118                 put_unaligned(found_key.offset, &entry->offset);
6119                 entries++;
6120                 addr += sizeof(struct dir_entry) + name_len;
6121                 total_len += sizeof(struct dir_entry) + name_len;
6122 next:
6123                 path->slots[0]++;
6124         }
6125         btrfs_release_path(path);
6126
6127         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6128         if (ret)
6129                 goto nopos;
6130
6131         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
6132         if (ret)
6133                 goto nopos;
6134
6135         /*
6136          * Stop new entries from being returned after we return the last
6137          * entry.
6138          *
6139          * New directory entries are assigned a strictly increasing
6140          * offset.  This means that new entries created during readdir
6141          * are *guaranteed* to be seen in the future by that readdir.
6142          * This has broken buggy programs which operate on names as
6143          * they're returned by readdir.  Until we re-use freed offsets
6144          * we have this hack to stop new entries from being returned
6145          * under the assumption that they'll never reach this huge
6146          * offset.
6147          *
6148          * This is being careful not to overflow 32bit loff_t unless the
6149          * last entry requires it because doing so has broken 32bit apps
6150          * in the past.
6151          */
6152         if (ctx->pos >= INT_MAX)
6153                 ctx->pos = LLONG_MAX;
6154         else
6155                 ctx->pos = INT_MAX;
6156 nopos:
6157         ret = 0;
6158 err:
6159         if (put)
6160                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
6161         btrfs_free_path(path);
6162         return ret;
6163 }
6164
6165 /*
6166  * This is somewhat expensive, updating the tree every time the
6167  * inode changes.  But, it is most likely to find the inode in cache.
6168  * FIXME, needs more benchmarking...there are no reasons other than performance
6169  * to keep or drop this code.
6170  */
6171 static int btrfs_dirty_inode(struct inode *inode)
6172 {
6173         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6174         struct btrfs_root *root = BTRFS_I(inode)->root;
6175         struct btrfs_trans_handle *trans;
6176         int ret;
6177
6178         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6179                 return 0;
6180
6181         trans = btrfs_join_transaction(root);
6182         if (IS_ERR(trans))
6183                 return PTR_ERR(trans);
6184
6185         ret = btrfs_update_inode(trans, root, inode);
6186         if (ret && ret == -ENOSPC) {
6187                 /* whoops, lets try again with the full transaction */
6188                 btrfs_end_transaction(trans);
6189                 trans = btrfs_start_transaction(root, 1);
6190                 if (IS_ERR(trans))
6191                         return PTR_ERR(trans);
6192
6193                 ret = btrfs_update_inode(trans, root, inode);
6194         }
6195         btrfs_end_transaction(trans);
6196         if (BTRFS_I(inode)->delayed_node)
6197                 btrfs_balance_delayed_items(fs_info);
6198
6199         return ret;
6200 }
6201
6202 /*
6203  * This is a copy of file_update_time.  We need this so we can return error on
6204  * ENOSPC for updating the inode in the case of file write and mmap writes.
6205  */
6206 static int btrfs_update_time(struct inode *inode, struct timespec *now,
6207                              int flags)
6208 {
6209         struct btrfs_root *root = BTRFS_I(inode)->root;
6210
6211         if (btrfs_root_readonly(root))
6212                 return -EROFS;
6213
6214         if (flags & S_VERSION)
6215                 inode_inc_iversion(inode);
6216         if (flags & S_CTIME)
6217                 inode->i_ctime = *now;
6218         if (flags & S_MTIME)
6219                 inode->i_mtime = *now;
6220         if (flags & S_ATIME)
6221                 inode->i_atime = *now;
6222         return btrfs_dirty_inode(inode);
6223 }
6224
6225 /*
6226  * find the highest existing sequence number in a directory
6227  * and then set the in-memory index_cnt variable to reflect
6228  * free sequence numbers
6229  */
6230 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6231 {
6232         struct btrfs_root *root = inode->root;
6233         struct btrfs_key key, found_key;
6234         struct btrfs_path *path;
6235         struct extent_buffer *leaf;
6236         int ret;
6237
6238         key.objectid = btrfs_ino(inode);
6239         key.type = BTRFS_DIR_INDEX_KEY;
6240         key.offset = (u64)-1;
6241
6242         path = btrfs_alloc_path();
6243         if (!path)
6244                 return -ENOMEM;
6245
6246         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6247         if (ret < 0)
6248                 goto out;
6249         /* FIXME: we should be able to handle this */
6250         if (ret == 0)
6251                 goto out;
6252         ret = 0;
6253
6254         /*
6255          * MAGIC NUMBER EXPLANATION:
6256          * since we search a directory based on f_pos we have to start at 2
6257          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6258          * else has to start at 2
6259          */
6260         if (path->slots[0] == 0) {
6261                 inode->index_cnt = 2;
6262                 goto out;
6263         }
6264
6265         path->slots[0]--;
6266
6267         leaf = path->nodes[0];
6268         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6269
6270         if (found_key.objectid != btrfs_ino(inode) ||
6271             found_key.type != BTRFS_DIR_INDEX_KEY) {
6272                 inode->index_cnt = 2;
6273                 goto out;
6274         }
6275
6276         inode->index_cnt = found_key.offset + 1;
6277 out:
6278         btrfs_free_path(path);
6279         return ret;
6280 }
6281
6282 /*
6283  * helper to find a free sequence number in a given directory.  This current
6284  * code is very simple, later versions will do smarter things in the btree
6285  */
6286 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6287 {
6288         int ret = 0;
6289
6290         if (dir->index_cnt == (u64)-1) {
6291                 ret = btrfs_inode_delayed_dir_index_count(dir);
6292                 if (ret) {
6293                         ret = btrfs_set_inode_index_count(dir);
6294                         if (ret)
6295                                 return ret;
6296                 }
6297         }
6298
6299         *index = dir->index_cnt;
6300         dir->index_cnt++;
6301
6302         return ret;
6303 }
6304
6305 static int btrfs_insert_inode_locked(struct inode *inode)
6306 {
6307         struct btrfs_iget_args args;
6308         args.location = &BTRFS_I(inode)->location;
6309         args.root = BTRFS_I(inode)->root;
6310
6311         return insert_inode_locked4(inode,
6312                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6313                    btrfs_find_actor, &args);
6314 }
6315
6316 /*
6317  * Inherit flags from the parent inode.
6318  *
6319  * Currently only the compression flags and the cow flags are inherited.
6320  */
6321 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6322 {
6323         unsigned int flags;
6324
6325         if (!dir)
6326                 return;
6327
6328         flags = BTRFS_I(dir)->flags;
6329
6330         if (flags & BTRFS_INODE_NOCOMPRESS) {
6331                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6332                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6333         } else if (flags & BTRFS_INODE_COMPRESS) {
6334                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6335                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6336         }
6337
6338         if (flags & BTRFS_INODE_NODATACOW) {
6339                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6340                 if (S_ISREG(inode->i_mode))
6341                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6342         }
6343
6344         btrfs_update_iflags(inode);
6345 }
6346
6347 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6348                                      struct btrfs_root *root,
6349                                      struct inode *dir,
6350                                      const char *name, int name_len,
6351                                      u64 ref_objectid, u64 objectid,
6352                                      umode_t mode, u64 *index)
6353 {
6354         struct btrfs_fs_info *fs_info = root->fs_info;
6355         struct inode *inode;
6356         struct btrfs_inode_item *inode_item;
6357         struct btrfs_key *location;
6358         struct btrfs_path *path;
6359         struct btrfs_inode_ref *ref;
6360         struct btrfs_key key[2];
6361         u32 sizes[2];
6362         int nitems = name ? 2 : 1;
6363         unsigned long ptr;
6364         int ret;
6365
6366         path = btrfs_alloc_path();
6367         if (!path)
6368                 return ERR_PTR(-ENOMEM);
6369
6370         inode = new_inode(fs_info->sb);
6371         if (!inode) {
6372                 btrfs_free_path(path);
6373                 return ERR_PTR(-ENOMEM);
6374         }
6375
6376         /*
6377          * O_TMPFILE, set link count to 0, so that after this point,
6378          * we fill in an inode item with the correct link count.
6379          */
6380         if (!name)
6381                 set_nlink(inode, 0);
6382
6383         /*
6384          * we have to initialize this early, so we can reclaim the inode
6385          * number if we fail afterwards in this function.
6386          */
6387         inode->i_ino = objectid;
6388
6389         if (dir && name) {
6390                 trace_btrfs_inode_request(dir);
6391
6392                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6393                 if (ret) {
6394                         btrfs_free_path(path);
6395                         iput(inode);
6396                         return ERR_PTR(ret);
6397                 }
6398         } else if (dir) {
6399                 *index = 0;
6400         }
6401         /*
6402          * index_cnt is ignored for everything but a dir,
6403          * btrfs_get_inode_index_count has an explanation for the magic
6404          * number
6405          */
6406         BTRFS_I(inode)->index_cnt = 2;
6407         BTRFS_I(inode)->dir_index = *index;
6408         BTRFS_I(inode)->root = root;
6409         BTRFS_I(inode)->generation = trans->transid;
6410         inode->i_generation = BTRFS_I(inode)->generation;
6411
6412         /*
6413          * We could have gotten an inode number from somebody who was fsynced
6414          * and then removed in this same transaction, so let's just set full
6415          * sync since it will be a full sync anyway and this will blow away the
6416          * old info in the log.
6417          */
6418         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6419
6420         key[0].objectid = objectid;
6421         key[0].type = BTRFS_INODE_ITEM_KEY;
6422         key[0].offset = 0;
6423
6424         sizes[0] = sizeof(struct btrfs_inode_item);
6425
6426         if (name) {
6427                 /*
6428                  * Start new inodes with an inode_ref. This is slightly more
6429                  * efficient for small numbers of hard links since they will
6430                  * be packed into one item. Extended refs will kick in if we
6431                  * add more hard links than can fit in the ref item.
6432                  */
6433                 key[1].objectid = objectid;
6434                 key[1].type = BTRFS_INODE_REF_KEY;
6435                 key[1].offset = ref_objectid;
6436
6437                 sizes[1] = name_len + sizeof(*ref);
6438         }
6439
6440         location = &BTRFS_I(inode)->location;
6441         location->objectid = objectid;
6442         location->offset = 0;
6443         location->type = BTRFS_INODE_ITEM_KEY;
6444
6445         ret = btrfs_insert_inode_locked(inode);
6446         if (ret < 0)
6447                 goto fail;
6448
6449         path->leave_spinning = 1;
6450         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6451         if (ret != 0)
6452                 goto fail_unlock;
6453
6454         inode_init_owner(inode, dir, mode);
6455         inode_set_bytes(inode, 0);
6456
6457         inode->i_mtime = current_time(inode);
6458         inode->i_atime = inode->i_mtime;
6459         inode->i_ctime = inode->i_mtime;
6460         BTRFS_I(inode)->i_otime = inode->i_mtime;
6461
6462         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6463                                   struct btrfs_inode_item);
6464         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6465                              sizeof(*inode_item));
6466         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6467
6468         if (name) {
6469                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6470                                      struct btrfs_inode_ref);
6471                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6472                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6473                 ptr = (unsigned long)(ref + 1);
6474                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6475         }
6476
6477         btrfs_mark_buffer_dirty(path->nodes[0]);
6478         btrfs_free_path(path);
6479
6480         btrfs_inherit_iflags(inode, dir);
6481
6482         if (S_ISREG(mode)) {
6483                 if (btrfs_test_opt(fs_info, NODATASUM))
6484                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6485                 if (btrfs_test_opt(fs_info, NODATACOW))
6486                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6487                                 BTRFS_INODE_NODATASUM;
6488         }
6489
6490         inode_tree_add(inode);
6491
6492         trace_btrfs_inode_new(inode);
6493         btrfs_set_inode_last_trans(trans, inode);
6494
6495         btrfs_update_root_times(trans, root);
6496
6497         ret = btrfs_inode_inherit_props(trans, inode, dir);
6498         if (ret)
6499                 btrfs_err(fs_info,
6500                           "error inheriting props for ino %llu (root %llu): %d",
6501                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6502
6503         return inode;
6504
6505 fail_unlock:
6506         unlock_new_inode(inode);
6507 fail:
6508         if (dir && name)
6509                 BTRFS_I(dir)->index_cnt--;
6510         btrfs_free_path(path);
6511         iput(inode);
6512         return ERR_PTR(ret);
6513 }
6514
6515 static inline u8 btrfs_inode_type(struct inode *inode)
6516 {
6517         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6518 }
6519
6520 /*
6521  * utility function to add 'inode' into 'parent_inode' with
6522  * a give name and a given sequence number.
6523  * if 'add_backref' is true, also insert a backref from the
6524  * inode to the parent directory.
6525  */
6526 int btrfs_add_link(struct btrfs_trans_handle *trans,
6527                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6528                    const char *name, int name_len, int add_backref, u64 index)
6529 {
6530         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6531         int ret = 0;
6532         struct btrfs_key key;
6533         struct btrfs_root *root = parent_inode->root;
6534         u64 ino = btrfs_ino(inode);
6535         u64 parent_ino = btrfs_ino(parent_inode);
6536
6537         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6538                 memcpy(&key, &inode->root->root_key, sizeof(key));
6539         } else {
6540                 key.objectid = ino;
6541                 key.type = BTRFS_INODE_ITEM_KEY;
6542                 key.offset = 0;
6543         }
6544
6545         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6546                 ret = btrfs_add_root_ref(trans, fs_info, key.objectid,
6547                                          root->root_key.objectid, parent_ino,
6548                                          index, name, name_len);
6549         } else if (add_backref) {
6550                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6551                                              parent_ino, index);
6552         }
6553
6554         /* Nothing to clean up yet */
6555         if (ret)
6556                 return ret;
6557
6558         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6559                                     parent_inode, &key,
6560                                     btrfs_inode_type(&inode->vfs_inode), index);
6561         if (ret == -EEXIST || ret == -EOVERFLOW)
6562                 goto fail_dir_item;
6563         else if (ret) {
6564                 btrfs_abort_transaction(trans, ret);
6565                 return ret;
6566         }
6567
6568         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6569                            name_len * 2);
6570         inode_inc_iversion(&parent_inode->vfs_inode);
6571         parent_inode->vfs_inode.i_mtime = parent_inode->vfs_inode.i_ctime =
6572                 current_time(&parent_inode->vfs_inode);
6573         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6574         if (ret)
6575                 btrfs_abort_transaction(trans, ret);
6576         return ret;
6577
6578 fail_dir_item:
6579         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6580                 u64 local_index;
6581                 int err;
6582                 err = btrfs_del_root_ref(trans, fs_info, key.objectid,
6583                                          root->root_key.objectid, parent_ino,
6584                                          &local_index, name, name_len);
6585
6586         } else if (add_backref) {
6587                 u64 local_index;
6588                 int err;
6589
6590                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6591                                           ino, parent_ino, &local_index);
6592         }
6593         return ret;
6594 }
6595
6596 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6597                             struct btrfs_inode *dir, struct dentry *dentry,
6598                             struct btrfs_inode *inode, int backref, u64 index)
6599 {
6600         int err = btrfs_add_link(trans, dir, inode,
6601                                  dentry->d_name.name, dentry->d_name.len,
6602                                  backref, index);
6603         if (err > 0)
6604                 err = -EEXIST;
6605         return err;
6606 }
6607
6608 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6609                         umode_t mode, dev_t rdev)
6610 {
6611         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6612         struct btrfs_trans_handle *trans;
6613         struct btrfs_root *root = BTRFS_I(dir)->root;
6614         struct inode *inode = NULL;
6615         int err;
6616         int drop_inode = 0;
6617         u64 objectid;
6618         u64 index = 0;
6619
6620         /*
6621          * 2 for inode item and ref
6622          * 2 for dir items
6623          * 1 for xattr if selinux is on
6624          */
6625         trans = btrfs_start_transaction(root, 5);
6626         if (IS_ERR(trans))
6627                 return PTR_ERR(trans);
6628
6629         err = btrfs_find_free_ino(root, &objectid);
6630         if (err)
6631                 goto out_unlock;
6632
6633         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6634                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6635                         mode, &index);
6636         if (IS_ERR(inode)) {
6637                 err = PTR_ERR(inode);
6638                 goto out_unlock;
6639         }
6640
6641         /*
6642         * If the active LSM wants to access the inode during
6643         * d_instantiate it needs these. Smack checks to see
6644         * if the filesystem supports xattrs by looking at the
6645         * ops vector.
6646         */
6647         inode->i_op = &btrfs_special_inode_operations;
6648         init_special_inode(inode, inode->i_mode, rdev);
6649
6650         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6651         if (err)
6652                 goto out_unlock_inode;
6653
6654         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6655                         0, index);
6656         if (err) {
6657                 goto out_unlock_inode;
6658         } else {
6659                 btrfs_update_inode(trans, root, inode);
6660                 d_instantiate_new(dentry, inode);
6661         }
6662
6663 out_unlock:
6664         btrfs_end_transaction(trans);
6665         btrfs_balance_delayed_items(fs_info);
6666         btrfs_btree_balance_dirty(fs_info);
6667         if (drop_inode) {
6668                 inode_dec_link_count(inode);
6669                 iput(inode);
6670         }
6671         return err;
6672
6673 out_unlock_inode:
6674         drop_inode = 1;
6675         unlock_new_inode(inode);
6676         goto out_unlock;
6677
6678 }
6679
6680 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6681                         umode_t mode, bool excl)
6682 {
6683         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6684         struct btrfs_trans_handle *trans;
6685         struct btrfs_root *root = BTRFS_I(dir)->root;
6686         struct inode *inode = NULL;
6687         int drop_inode_on_err = 0;
6688         int err;
6689         u64 objectid;
6690         u64 index = 0;
6691
6692         /*
6693          * 2 for inode item and ref
6694          * 2 for dir items
6695          * 1 for xattr if selinux is on
6696          */
6697         trans = btrfs_start_transaction(root, 5);
6698         if (IS_ERR(trans))
6699                 return PTR_ERR(trans);
6700
6701         err = btrfs_find_free_ino(root, &objectid);
6702         if (err)
6703                 goto out_unlock;
6704
6705         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6706                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6707                         mode, &index);
6708         if (IS_ERR(inode)) {
6709                 err = PTR_ERR(inode);
6710                 goto out_unlock;
6711         }
6712         drop_inode_on_err = 1;
6713         /*
6714         * If the active LSM wants to access the inode during
6715         * d_instantiate it needs these. Smack checks to see
6716         * if the filesystem supports xattrs by looking at the
6717         * ops vector.
6718         */
6719         inode->i_fop = &btrfs_file_operations;
6720         inode->i_op = &btrfs_file_inode_operations;
6721         inode->i_mapping->a_ops = &btrfs_aops;
6722
6723         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6724         if (err)
6725                 goto out_unlock_inode;
6726
6727         err = btrfs_update_inode(trans, root, inode);
6728         if (err)
6729                 goto out_unlock_inode;
6730
6731         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6732                         0, index);
6733         if (err)
6734                 goto out_unlock_inode;
6735
6736         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6737         d_instantiate_new(dentry, inode);
6738
6739 out_unlock:
6740         btrfs_end_transaction(trans);
6741         if (err && drop_inode_on_err) {
6742                 inode_dec_link_count(inode);
6743                 iput(inode);
6744         }
6745         btrfs_balance_delayed_items(fs_info);
6746         btrfs_btree_balance_dirty(fs_info);
6747         return err;
6748
6749 out_unlock_inode:
6750         unlock_new_inode(inode);
6751         goto out_unlock;
6752
6753 }
6754
6755 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6756                       struct dentry *dentry)
6757 {
6758         struct btrfs_trans_handle *trans = NULL;
6759         struct btrfs_root *root = BTRFS_I(dir)->root;
6760         struct inode *inode = d_inode(old_dentry);
6761         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6762         u64 index;
6763         int err;
6764         int drop_inode = 0;
6765
6766         /* do not allow sys_link's with other subvols of the same device */
6767         if (root->objectid != BTRFS_I(inode)->root->objectid)
6768                 return -EXDEV;
6769
6770         if (inode->i_nlink >= BTRFS_LINK_MAX)
6771                 return -EMLINK;
6772
6773         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6774         if (err)
6775                 goto fail;
6776
6777         /*
6778          * 2 items for inode and inode ref
6779          * 2 items for dir items
6780          * 1 item for parent inode
6781          */
6782         trans = btrfs_start_transaction(root, 5);
6783         if (IS_ERR(trans)) {
6784                 err = PTR_ERR(trans);
6785                 trans = NULL;
6786                 goto fail;
6787         }
6788
6789         /* There are several dir indexes for this inode, clear the cache. */
6790         BTRFS_I(inode)->dir_index = 0ULL;
6791         inc_nlink(inode);
6792         inode_inc_iversion(inode);
6793         inode->i_ctime = current_time(inode);
6794         ihold(inode);
6795         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6796
6797         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6798                         1, index);
6799
6800         if (err) {
6801                 drop_inode = 1;
6802         } else {
6803                 struct dentry *parent = dentry->d_parent;
6804                 err = btrfs_update_inode(trans, root, inode);
6805                 if (err)
6806                         goto fail;
6807                 if (inode->i_nlink == 1) {
6808                         /*
6809                          * If new hard link count is 1, it's a file created
6810                          * with open(2) O_TMPFILE flag.
6811                          */
6812                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6813                         if (err)
6814                                 goto fail;
6815                 }
6816                 d_instantiate(dentry, inode);
6817                 btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent);
6818         }
6819
6820         btrfs_balance_delayed_items(fs_info);
6821 fail:
6822         if (trans)
6823                 btrfs_end_transaction(trans);
6824         if (drop_inode) {
6825                 inode_dec_link_count(inode);
6826                 iput(inode);
6827         }
6828         btrfs_btree_balance_dirty(fs_info);
6829         return err;
6830 }
6831
6832 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6833 {
6834         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6835         struct inode *inode = NULL;
6836         struct btrfs_trans_handle *trans;
6837         struct btrfs_root *root = BTRFS_I(dir)->root;
6838         int err = 0;
6839         int drop_on_err = 0;
6840         u64 objectid = 0;
6841         u64 index = 0;
6842
6843         /*
6844          * 2 items for inode and ref
6845          * 2 items for dir items
6846          * 1 for xattr if selinux is on
6847          */
6848         trans = btrfs_start_transaction(root, 5);
6849         if (IS_ERR(trans))
6850                 return PTR_ERR(trans);
6851
6852         err = btrfs_find_free_ino(root, &objectid);
6853         if (err)
6854                 goto out_fail;
6855
6856         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6857                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6858                         S_IFDIR | mode, &index);
6859         if (IS_ERR(inode)) {
6860                 err = PTR_ERR(inode);
6861                 goto out_fail;
6862         }
6863
6864         drop_on_err = 1;
6865         /* these must be set before we unlock the inode */
6866         inode->i_op = &btrfs_dir_inode_operations;
6867         inode->i_fop = &btrfs_dir_file_operations;
6868
6869         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6870         if (err)
6871                 goto out_fail_inode;
6872
6873         btrfs_i_size_write(BTRFS_I(inode), 0);
6874         err = btrfs_update_inode(trans, root, inode);
6875         if (err)
6876                 goto out_fail_inode;
6877
6878         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6879                         dentry->d_name.name,
6880                         dentry->d_name.len, 0, index);
6881         if (err)
6882                 goto out_fail_inode;
6883
6884         d_instantiate_new(dentry, inode);
6885         drop_on_err = 0;
6886
6887 out_fail:
6888         btrfs_end_transaction(trans);
6889         if (drop_on_err) {
6890                 inode_dec_link_count(inode);
6891                 iput(inode);
6892         }
6893         btrfs_balance_delayed_items(fs_info);
6894         btrfs_btree_balance_dirty(fs_info);
6895         return err;
6896
6897 out_fail_inode:
6898         unlock_new_inode(inode);
6899         goto out_fail;
6900 }
6901
6902 /* Find next extent map of a given extent map, caller needs to ensure locks */
6903 static struct extent_map *next_extent_map(struct extent_map *em)
6904 {
6905         struct rb_node *next;
6906
6907         next = rb_next(&em->rb_node);
6908         if (!next)
6909                 return NULL;
6910         return container_of(next, struct extent_map, rb_node);
6911 }
6912
6913 static struct extent_map *prev_extent_map(struct extent_map *em)
6914 {
6915         struct rb_node *prev;
6916
6917         prev = rb_prev(&em->rb_node);
6918         if (!prev)
6919                 return NULL;
6920         return container_of(prev, struct extent_map, rb_node);
6921 }
6922
6923 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6924  * the existing extent is the nearest extent to map_start,
6925  * and an extent that you want to insert, deal with overlap and insert
6926  * the best fitted new extent into the tree.
6927  */
6928 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6929                                 struct extent_map *existing,
6930                                 struct extent_map *em,
6931                                 u64 map_start)
6932 {
6933         struct extent_map *prev;
6934         struct extent_map *next;
6935         u64 start;
6936         u64 end;
6937         u64 start_diff;
6938
6939         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6940
6941         if (existing->start > map_start) {
6942                 next = existing;
6943                 prev = prev_extent_map(next);
6944         } else {
6945                 prev = existing;
6946                 next = next_extent_map(prev);
6947         }
6948
6949         start = prev ? extent_map_end(prev) : em->start;
6950         start = max_t(u64, start, em->start);
6951         end = next ? next->start : extent_map_end(em);
6952         end = min_t(u64, end, extent_map_end(em));
6953         start_diff = start - em->start;
6954         em->start = start;
6955         em->len = end - start;
6956         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6957             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6958                 em->block_start += start_diff;
6959                 em->block_len -= start_diff;
6960         }
6961         return add_extent_mapping(em_tree, em, 0);
6962 }
6963
6964 static noinline int uncompress_inline(struct btrfs_path *path,
6965                                       struct page *page,
6966                                       size_t pg_offset, u64 extent_offset,
6967                                       struct btrfs_file_extent_item *item)
6968 {
6969         int ret;
6970         struct extent_buffer *leaf = path->nodes[0];
6971         char *tmp;
6972         size_t max_size;
6973         unsigned long inline_size;
6974         unsigned long ptr;
6975         int compress_type;
6976
6977         WARN_ON(pg_offset != 0);
6978         compress_type = btrfs_file_extent_compression(leaf, item);
6979         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6980         inline_size = btrfs_file_extent_inline_item_len(leaf,
6981                                         btrfs_item_nr(path->slots[0]));
6982         tmp = kmalloc(inline_size, GFP_NOFS);
6983         if (!tmp)
6984                 return -ENOMEM;
6985         ptr = btrfs_file_extent_inline_start(item);
6986
6987         read_extent_buffer(leaf, tmp, ptr, inline_size);
6988
6989         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6990         ret = btrfs_decompress(compress_type, tmp, page,
6991                                extent_offset, inline_size, max_size);
6992
6993         /*
6994          * decompression code contains a memset to fill in any space between the end
6995          * of the uncompressed data and the end of max_size in case the decompressed
6996          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6997          * the end of an inline extent and the beginning of the next block, so we
6998          * cover that region here.
6999          */
7000
7001         if (max_size + pg_offset < PAGE_SIZE) {
7002                 char *map = kmap(page);
7003                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
7004                 kunmap(page);
7005         }
7006         kfree(tmp);
7007         return ret;
7008 }
7009
7010 /*
7011  * a bit scary, this does extent mapping from logical file offset to the disk.
7012  * the ugly parts come from merging extents from the disk with the in-ram
7013  * representation.  This gets more complex because of the data=ordered code,
7014  * where the in-ram extents might be locked pending data=ordered completion.
7015  *
7016  * This also copies inline extents directly into the page.
7017  */
7018 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
7019                 struct page *page,
7020             size_t pg_offset, u64 start, u64 len,
7021                 int create)
7022 {
7023         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
7024         int ret;
7025         int err = 0;
7026         u64 extent_start = 0;
7027         u64 extent_end = 0;
7028         u64 objectid = btrfs_ino(inode);
7029         u32 found_type;
7030         struct btrfs_path *path = NULL;
7031         struct btrfs_root *root = inode->root;
7032         struct btrfs_file_extent_item *item;
7033         struct extent_buffer *leaf;
7034         struct btrfs_key found_key;
7035         struct extent_map *em = NULL;
7036         struct extent_map_tree *em_tree = &inode->extent_tree;
7037         struct extent_io_tree *io_tree = &inode->io_tree;
7038         struct btrfs_trans_handle *trans = NULL;
7039         const bool new_inline = !page || create;
7040
7041 again:
7042         read_lock(&em_tree->lock);
7043         em = lookup_extent_mapping(em_tree, start, len);
7044         if (em)
7045                 em->bdev = fs_info->fs_devices->latest_bdev;
7046         read_unlock(&em_tree->lock);
7047
7048         if (em) {
7049                 if (em->start > start || em->start + em->len <= start)
7050                         free_extent_map(em);
7051                 else if (em->block_start == EXTENT_MAP_INLINE && page)
7052                         free_extent_map(em);
7053                 else
7054                         goto out;
7055         }
7056         em = alloc_extent_map();
7057         if (!em) {
7058                 err = -ENOMEM;
7059                 goto out;
7060         }
7061         em->bdev = fs_info->fs_devices->latest_bdev;
7062         em->start = EXTENT_MAP_HOLE;
7063         em->orig_start = EXTENT_MAP_HOLE;
7064         em->len = (u64)-1;
7065         em->block_len = (u64)-1;
7066
7067         if (!path) {
7068                 path = btrfs_alloc_path();
7069                 if (!path) {
7070                         err = -ENOMEM;
7071                         goto out;
7072                 }
7073                 /*
7074                  * Chances are we'll be called again, so go ahead and do
7075                  * readahead
7076                  */
7077                 path->reada = READA_FORWARD;
7078         }
7079
7080         ret = btrfs_lookup_file_extent(trans, root, path,
7081                                        objectid, start, trans != NULL);
7082         if (ret < 0) {
7083                 err = ret;
7084                 goto out;
7085         }
7086
7087         if (ret != 0) {
7088                 if (path->slots[0] == 0)
7089                         goto not_found;
7090                 path->slots[0]--;
7091         }
7092
7093         leaf = path->nodes[0];
7094         item = btrfs_item_ptr(leaf, path->slots[0],
7095                               struct btrfs_file_extent_item);
7096         /* are we inside the extent that was found? */
7097         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7098         found_type = found_key.type;
7099         if (found_key.objectid != objectid ||
7100             found_type != BTRFS_EXTENT_DATA_KEY) {
7101                 /*
7102                  * If we backup past the first extent we want to move forward
7103                  * and see if there is an extent in front of us, otherwise we'll
7104                  * say there is a hole for our whole search range which can
7105                  * cause problems.
7106                  */
7107                 extent_end = start;
7108                 goto next;
7109         }
7110
7111         found_type = btrfs_file_extent_type(leaf, item);
7112         extent_start = found_key.offset;
7113         if (found_type == BTRFS_FILE_EXTENT_REG ||
7114             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7115                 extent_end = extent_start +
7116                        btrfs_file_extent_num_bytes(leaf, item);
7117
7118                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
7119                                                        extent_start);
7120         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
7121                 size_t size;
7122                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
7123                 extent_end = ALIGN(extent_start + size,
7124                                    fs_info->sectorsize);
7125
7126                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
7127                                                       path->slots[0],
7128                                                       extent_start);
7129         }
7130 next:
7131         if (start >= extent_end) {
7132                 path->slots[0]++;
7133                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
7134                         ret = btrfs_next_leaf(root, path);
7135                         if (ret < 0) {
7136                                 err = ret;
7137                                 goto out;
7138                         }
7139                         if (ret > 0)
7140                                 goto not_found;
7141                         leaf = path->nodes[0];
7142                 }
7143                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7144                 if (found_key.objectid != objectid ||
7145                     found_key.type != BTRFS_EXTENT_DATA_KEY)
7146                         goto not_found;
7147                 if (start + len <= found_key.offset)
7148                         goto not_found;
7149                 if (start > found_key.offset)
7150                         goto next;
7151                 em->start = start;
7152                 em->orig_start = start;
7153                 em->len = found_key.offset - start;
7154                 goto not_found_em;
7155         }
7156
7157         btrfs_extent_item_to_extent_map(inode, path, item,
7158                         new_inline, em);
7159
7160         if (found_type == BTRFS_FILE_EXTENT_REG ||
7161             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7162                 goto insert;
7163         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
7164                 unsigned long ptr;
7165                 char *map;
7166                 size_t size;
7167                 size_t extent_offset;
7168                 size_t copy_size;
7169
7170                 if (new_inline)
7171                         goto out;
7172
7173                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
7174                 extent_offset = page_offset(page) + pg_offset - extent_start;
7175                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
7176                                   size - extent_offset);
7177                 em->start = extent_start + extent_offset;
7178                 em->len = ALIGN(copy_size, fs_info->sectorsize);
7179                 em->orig_block_len = em->len;
7180                 em->orig_start = em->start;
7181                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
7182                 if (create == 0 && !PageUptodate(page)) {
7183                         if (btrfs_file_extent_compression(leaf, item) !=
7184                             BTRFS_COMPRESS_NONE) {
7185                                 ret = uncompress_inline(path, page, pg_offset,
7186                                                         extent_offset, item);
7187                                 if (ret) {
7188                                         err = ret;
7189                                         goto out;
7190                                 }
7191                         } else {
7192                                 map = kmap(page);
7193                                 read_extent_buffer(leaf, map + pg_offset, ptr,
7194                                                    copy_size);
7195                                 if (pg_offset + copy_size < PAGE_SIZE) {
7196                                         memset(map + pg_offset + copy_size, 0,
7197                                                PAGE_SIZE - pg_offset -
7198                                                copy_size);
7199                                 }
7200                                 kunmap(page);
7201                         }
7202                         flush_dcache_page(page);
7203                 } else if (create && PageUptodate(page)) {
7204                         BUG();
7205                         if (!trans) {
7206                                 kunmap(page);
7207                                 free_extent_map(em);
7208                                 em = NULL;
7209
7210                                 btrfs_release_path(path);
7211                                 trans = btrfs_join_transaction(root);
7212
7213                                 if (IS_ERR(trans))
7214                                         return ERR_CAST(trans);
7215                                 goto again;
7216                         }
7217                         map = kmap(page);
7218                         write_extent_buffer(leaf, map + pg_offset, ptr,
7219                                             copy_size);
7220                         kunmap(page);
7221                         btrfs_mark_buffer_dirty(leaf);
7222                 }
7223                 set_extent_uptodate(io_tree, em->start,
7224                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7225                 goto insert;
7226         }
7227 not_found:
7228         em->start = start;
7229         em->orig_start = start;
7230         em->len = len;
7231 not_found_em:
7232         em->block_start = EXTENT_MAP_HOLE;
7233         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
7234 insert:
7235         btrfs_release_path(path);
7236         if (em->start > start || extent_map_end(em) <= start) {
7237                 btrfs_err(fs_info,
7238                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
7239                           em->start, em->len, start, len);
7240                 err = -EIO;
7241                 goto out;
7242         }
7243
7244         err = 0;
7245         write_lock(&em_tree->lock);
7246         ret = add_extent_mapping(em_tree, em, 0);
7247         /* it is possible that someone inserted the extent into the tree
7248          * while we had the lock dropped.  It is also possible that
7249          * an overlapping map exists in the tree
7250          */
7251         if (ret == -EEXIST) {
7252                 struct extent_map *existing;
7253
7254                 ret = 0;
7255
7256                 existing = search_extent_mapping(em_tree, start, len);
7257                 /*
7258                  * existing will always be non-NULL, since there must be
7259                  * extent causing the -EEXIST.
7260                  */
7261                 if (start >= existing->start &&
7262                     start < extent_map_end(existing)) {
7263                         free_extent_map(em);
7264                         em = existing;
7265                         err = 0;
7266                 } else {
7267                         /*
7268                          * The existing extent map is the one nearest to
7269                          * the [start, start + len) range which overlaps
7270                          */
7271                         err = merge_extent_mapping(em_tree, existing,
7272                                                    em, start);
7273                         free_extent_map(existing);
7274                         if (err) {
7275                                 free_extent_map(em);
7276                                 em = NULL;
7277                         }
7278                 }
7279         }
7280         write_unlock(&em_tree->lock);
7281 out:
7282
7283         trace_btrfs_get_extent(root, inode, em);
7284
7285         btrfs_free_path(path);
7286         if (trans) {
7287                 ret = btrfs_end_transaction(trans);
7288                 if (!err)
7289                         err = ret;
7290         }
7291         if (err) {
7292                 free_extent_map(em);
7293                 return ERR_PTR(err);
7294         }
7295         BUG_ON(!em); /* Error is always set */
7296         return em;
7297 }
7298
7299 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7300                 struct page *page,
7301                 size_t pg_offset, u64 start, u64 len,
7302                 int create)
7303 {
7304         struct extent_map *em;
7305         struct extent_map *hole_em = NULL;
7306         u64 range_start = start;
7307         u64 end;
7308         u64 found;
7309         u64 found_end;
7310         int err = 0;
7311
7312         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7313         if (IS_ERR(em))
7314                 return em;
7315         /*
7316          * If our em maps to:
7317          * - a hole or
7318          * - a pre-alloc extent,
7319          * there might actually be delalloc bytes behind it.
7320          */
7321         if (em->block_start != EXTENT_MAP_HOLE &&
7322             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7323                 return em;
7324         else
7325                 hole_em = em;
7326
7327         /* check to see if we've wrapped (len == -1 or similar) */
7328         end = start + len;
7329         if (end < start)
7330                 end = (u64)-1;
7331         else
7332                 end -= 1;
7333
7334         em = NULL;
7335
7336         /* ok, we didn't find anything, lets look for delalloc */
7337         found = count_range_bits(&inode->io_tree, &range_start,
7338                                  end, len, EXTENT_DELALLOC, 1);
7339         found_end = range_start + found;
7340         if (found_end < range_start)
7341                 found_end = (u64)-1;
7342
7343         /*
7344          * we didn't find anything useful, return
7345          * the original results from get_extent()
7346          */
7347         if (range_start > end || found_end <= start) {
7348                 em = hole_em;
7349                 hole_em = NULL;
7350                 goto out;
7351         }
7352
7353         /* adjust the range_start to make sure it doesn't
7354          * go backwards from the start they passed in
7355          */
7356         range_start = max(start, range_start);
7357         found = found_end - range_start;
7358
7359         if (found > 0) {
7360                 u64 hole_start = start;
7361                 u64 hole_len = len;
7362
7363                 em = alloc_extent_map();
7364                 if (!em) {
7365                         err = -ENOMEM;
7366                         goto out;
7367                 }
7368                 /*
7369                  * when btrfs_get_extent can't find anything it
7370                  * returns one huge hole
7371                  *
7372                  * make sure what it found really fits our range, and
7373                  * adjust to make sure it is based on the start from
7374                  * the caller
7375                  */
7376                 if (hole_em) {
7377                         u64 calc_end = extent_map_end(hole_em);
7378
7379                         if (calc_end <= start || (hole_em->start > end)) {
7380                                 free_extent_map(hole_em);
7381                                 hole_em = NULL;
7382                         } else {
7383                                 hole_start = max(hole_em->start, start);
7384                                 hole_len = calc_end - hole_start;
7385                         }
7386                 }
7387                 em->bdev = NULL;
7388                 if (hole_em && range_start > hole_start) {
7389                         /* our hole starts before our delalloc, so we
7390                          * have to return just the parts of the hole
7391                          * that go until  the delalloc starts
7392                          */
7393                         em->len = min(hole_len,
7394                                       range_start - hole_start);
7395                         em->start = hole_start;
7396                         em->orig_start = hole_start;
7397                         /*
7398                          * don't adjust block start at all,
7399                          * it is fixed at EXTENT_MAP_HOLE
7400                          */
7401                         em->block_start = hole_em->block_start;
7402                         em->block_len = hole_len;
7403                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7404                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7405                 } else {
7406                         em->start = range_start;
7407                         em->len = found;
7408                         em->orig_start = range_start;
7409                         em->block_start = EXTENT_MAP_DELALLOC;
7410                         em->block_len = found;
7411                 }
7412         } else if (hole_em) {
7413                 return hole_em;
7414         }
7415 out:
7416
7417         free_extent_map(hole_em);
7418         if (err) {
7419                 free_extent_map(em);
7420                 return ERR_PTR(err);
7421         }
7422         return em;
7423 }
7424
7425 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7426                                                   const u64 start,
7427                                                   const u64 len,
7428                                                   const u64 orig_start,
7429                                                   const u64 block_start,
7430                                                   const u64 block_len,
7431                                                   const u64 orig_block_len,
7432                                                   const u64 ram_bytes,
7433                                                   const int type)
7434 {
7435         struct extent_map *em = NULL;
7436         int ret;
7437
7438         if (type != BTRFS_ORDERED_NOCOW) {
7439                 em = create_io_em(inode, start, len, orig_start,
7440                                   block_start, block_len, orig_block_len,
7441                                   ram_bytes,
7442                                   BTRFS_COMPRESS_NONE, /* compress_type */
7443                                   type);
7444                 if (IS_ERR(em))
7445                         goto out;
7446         }
7447         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7448                                            len, block_len, type);
7449         if (ret) {
7450                 if (em) {
7451                         free_extent_map(em);
7452                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7453                                                 start + len - 1, 0);
7454                 }
7455                 em = ERR_PTR(ret);
7456         }
7457  out:
7458
7459         return em;
7460 }
7461
7462 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7463                                                   u64 start, u64 len)
7464 {
7465         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7466         struct btrfs_root *root = BTRFS_I(inode)->root;
7467         struct extent_map *em;
7468         struct btrfs_key ins;
7469         u64 alloc_hint;
7470         int ret;
7471
7472         alloc_hint = get_extent_allocation_hint(inode, start, len);
7473         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7474                                    0, alloc_hint, &ins, 1, 1);
7475         if (ret)
7476                 return ERR_PTR(ret);
7477
7478         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7479                                      ins.objectid, ins.offset, ins.offset,
7480                                      ins.offset, BTRFS_ORDERED_REGULAR);
7481         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7482         if (IS_ERR(em))
7483                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7484                                            ins.offset, 1);
7485
7486         return em;
7487 }
7488
7489 /*
7490  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7491  * block must be cow'd
7492  */
7493 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7494                               u64 *orig_start, u64 *orig_block_len,
7495                               u64 *ram_bytes)
7496 {
7497         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7498         struct btrfs_path *path;
7499         int ret;
7500         struct extent_buffer *leaf;
7501         struct btrfs_root *root = BTRFS_I(inode)->root;
7502         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7503         struct btrfs_file_extent_item *fi;
7504         struct btrfs_key key;
7505         u64 disk_bytenr;
7506         u64 backref_offset;
7507         u64 extent_end;
7508         u64 num_bytes;
7509         int slot;
7510         int found_type;
7511         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7512
7513         path = btrfs_alloc_path();
7514         if (!path)
7515                 return -ENOMEM;
7516
7517         ret = btrfs_lookup_file_extent(NULL, root, path,
7518                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7519         if (ret < 0)
7520                 goto out;
7521
7522         slot = path->slots[0];
7523         if (ret == 1) {
7524                 if (slot == 0) {
7525                         /* can't find the item, must cow */
7526                         ret = 0;
7527                         goto out;
7528                 }
7529                 slot--;
7530         }
7531         ret = 0;
7532         leaf = path->nodes[0];
7533         btrfs_item_key_to_cpu(leaf, &key, slot);
7534         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7535             key.type != BTRFS_EXTENT_DATA_KEY) {
7536                 /* not our file or wrong item type, must cow */
7537                 goto out;
7538         }
7539
7540         if (key.offset > offset) {
7541                 /* Wrong offset, must cow */
7542                 goto out;
7543         }
7544
7545         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7546         found_type = btrfs_file_extent_type(leaf, fi);
7547         if (found_type != BTRFS_FILE_EXTENT_REG &&
7548             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7549                 /* not a regular extent, must cow */
7550                 goto out;
7551         }
7552
7553         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7554                 goto out;
7555
7556         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7557         if (extent_end <= offset)
7558                 goto out;
7559
7560         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7561         if (disk_bytenr == 0)
7562                 goto out;
7563
7564         if (btrfs_file_extent_compression(leaf, fi) ||
7565             btrfs_file_extent_encryption(leaf, fi) ||
7566             btrfs_file_extent_other_encoding(leaf, fi))
7567                 goto out;
7568
7569         backref_offset = btrfs_file_extent_offset(leaf, fi);
7570
7571         if (orig_start) {
7572                 *orig_start = key.offset - backref_offset;
7573                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7574                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7575         }
7576
7577         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7578                 goto out;
7579
7580         num_bytes = min(offset + *len, extent_end) - offset;
7581         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7582                 u64 range_end;
7583
7584                 range_end = round_up(offset + num_bytes,
7585                                      root->fs_info->sectorsize) - 1;
7586                 ret = test_range_bit(io_tree, offset, range_end,
7587                                      EXTENT_DELALLOC, 0, NULL);
7588                 if (ret) {
7589                         ret = -EAGAIN;
7590                         goto out;
7591                 }
7592         }
7593
7594         btrfs_release_path(path);
7595
7596         /*
7597          * look for other files referencing this extent, if we
7598          * find any we must cow
7599          */
7600
7601         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7602                                     key.offset - backref_offset, disk_bytenr);
7603         if (ret) {
7604                 ret = 0;
7605                 goto out;
7606         }
7607
7608         /*
7609          * adjust disk_bytenr and num_bytes to cover just the bytes
7610          * in this extent we are about to write.  If there
7611          * are any csums in that range we have to cow in order
7612          * to keep the csums correct
7613          */
7614         disk_bytenr += backref_offset;
7615         disk_bytenr += offset - key.offset;
7616         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7617                 goto out;
7618         /*
7619          * all of the above have passed, it is safe to overwrite this extent
7620          * without cow
7621          */
7622         *len = num_bytes;
7623         ret = 1;
7624 out:
7625         btrfs_free_path(path);
7626         return ret;
7627 }
7628
7629 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7630 {
7631         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7632         bool found = false;
7633         void **pagep = NULL;
7634         struct page *page = NULL;
7635         unsigned long start_idx;
7636         unsigned long end_idx;
7637
7638         start_idx = start >> PAGE_SHIFT;
7639
7640         /*
7641          * end is the last byte in the last page.  end == start is legal
7642          */
7643         end_idx = end >> PAGE_SHIFT;
7644
7645         rcu_read_lock();
7646
7647         /* Most of the code in this while loop is lifted from
7648          * find_get_page.  It's been modified to begin searching from a
7649          * page and return just the first page found in that range.  If the
7650          * found idx is less than or equal to the end idx then we know that
7651          * a page exists.  If no pages are found or if those pages are
7652          * outside of the range then we're fine (yay!) */
7653         while (page == NULL &&
7654                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7655                 page = radix_tree_deref_slot(pagep);
7656                 if (unlikely(!page))
7657                         break;
7658
7659                 if (radix_tree_exception(page)) {
7660                         if (radix_tree_deref_retry(page)) {
7661                                 page = NULL;
7662                                 continue;
7663                         }
7664                         /*
7665                          * Otherwise, shmem/tmpfs must be storing a swap entry
7666                          * here as an exceptional entry: so return it without
7667                          * attempting to raise page count.
7668                          */
7669                         page = NULL;
7670                         break; /* TODO: Is this relevant for this use case? */
7671                 }
7672
7673                 if (!page_cache_get_speculative(page)) {
7674                         page = NULL;
7675                         continue;
7676                 }
7677
7678                 /*
7679                  * Has the page moved?
7680                  * This is part of the lockless pagecache protocol. See
7681                  * include/linux/pagemap.h for details.
7682                  */
7683                 if (unlikely(page != *pagep)) {
7684                         put_page(page);
7685                         page = NULL;
7686                 }
7687         }
7688
7689         if (page) {
7690                 if (page->index <= end_idx)
7691                         found = true;
7692                 put_page(page);
7693         }
7694
7695         rcu_read_unlock();
7696         return found;
7697 }
7698
7699 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7700                               struct extent_state **cached_state, int writing)
7701 {
7702         struct btrfs_ordered_extent *ordered;
7703         int ret = 0;
7704
7705         while (1) {
7706                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7707                                  cached_state);
7708                 /*
7709                  * We're concerned with the entire range that we're going to be
7710                  * doing DIO to, so we need to make sure there's no ordered
7711                  * extents in this range.
7712                  */
7713                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7714                                                      lockend - lockstart + 1);
7715
7716                 /*
7717                  * We need to make sure there are no buffered pages in this
7718                  * range either, we could have raced between the invalidate in
7719                  * generic_file_direct_write and locking the extent.  The
7720                  * invalidate needs to happen so that reads after a write do not
7721                  * get stale data.
7722                  */
7723                 if (!ordered &&
7724                     (!writing ||
7725                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7726                         break;
7727
7728                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7729                                      cached_state, GFP_NOFS);
7730
7731                 if (ordered) {
7732                         /*
7733                          * If we are doing a DIO read and the ordered extent we
7734                          * found is for a buffered write, we can not wait for it
7735                          * to complete and retry, because if we do so we can
7736                          * deadlock with concurrent buffered writes on page
7737                          * locks. This happens only if our DIO read covers more
7738                          * than one extent map, if at this point has already
7739                          * created an ordered extent for a previous extent map
7740                          * and locked its range in the inode's io tree, and a
7741                          * concurrent write against that previous extent map's
7742                          * range and this range started (we unlock the ranges
7743                          * in the io tree only when the bios complete and
7744                          * buffered writes always lock pages before attempting
7745                          * to lock range in the io tree).
7746                          */
7747                         if (writing ||
7748                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7749                                 btrfs_start_ordered_extent(inode, ordered, 1);
7750                         else
7751                                 ret = -ENOTBLK;
7752                         btrfs_put_ordered_extent(ordered);
7753                 } else {
7754                         /*
7755                          * We could trigger writeback for this range (and wait
7756                          * for it to complete) and then invalidate the pages for
7757                          * this range (through invalidate_inode_pages2_range()),
7758                          * but that can lead us to a deadlock with a concurrent
7759                          * call to readpages() (a buffered read or a defrag call
7760                          * triggered a readahead) on a page lock due to an
7761                          * ordered dio extent we created before but did not have
7762                          * yet a corresponding bio submitted (whence it can not
7763                          * complete), which makes readpages() wait for that
7764                          * ordered extent to complete while holding a lock on
7765                          * that page.
7766                          */
7767                         ret = -ENOTBLK;
7768                 }
7769
7770                 if (ret)
7771                         break;
7772
7773                 cond_resched();
7774         }
7775
7776         return ret;
7777 }
7778
7779 /* The callers of this must take lock_extent() */
7780 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7781                                        u64 orig_start, u64 block_start,
7782                                        u64 block_len, u64 orig_block_len,
7783                                        u64 ram_bytes, int compress_type,
7784                                        int type)
7785 {
7786         struct extent_map_tree *em_tree;
7787         struct extent_map *em;
7788         struct btrfs_root *root = BTRFS_I(inode)->root;
7789         int ret;
7790
7791         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7792                type == BTRFS_ORDERED_COMPRESSED ||
7793                type == BTRFS_ORDERED_NOCOW ||
7794                type == BTRFS_ORDERED_REGULAR);
7795
7796         em_tree = &BTRFS_I(inode)->extent_tree;
7797         em = alloc_extent_map();
7798         if (!em)
7799                 return ERR_PTR(-ENOMEM);
7800
7801         em->start = start;
7802         em->orig_start = orig_start;
7803         em->len = len;
7804         em->block_len = block_len;
7805         em->block_start = block_start;
7806         em->bdev = root->fs_info->fs_devices->latest_bdev;
7807         em->orig_block_len = orig_block_len;
7808         em->ram_bytes = ram_bytes;
7809         em->generation = -1;
7810         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7811         if (type == BTRFS_ORDERED_PREALLOC) {
7812                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7813         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7814                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7815                 em->compress_type = compress_type;
7816         }
7817
7818         do {
7819                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7820                                 em->start + em->len - 1, 0);
7821                 write_lock(&em_tree->lock);
7822                 ret = add_extent_mapping(em_tree, em, 1);
7823                 write_unlock(&em_tree->lock);
7824                 /*
7825                  * The caller has taken lock_extent(), who could race with us
7826                  * to add em?
7827                  */
7828         } while (ret == -EEXIST);
7829
7830         if (ret) {
7831                 free_extent_map(em);
7832                 return ERR_PTR(ret);
7833         }
7834
7835         /* em got 2 refs now, callers needs to do free_extent_map once. */
7836         return em;
7837 }
7838
7839 static void adjust_dio_outstanding_extents(struct inode *inode,
7840                                            struct btrfs_dio_data *dio_data,
7841                                            const u64 len)
7842 {
7843         unsigned num_extents = count_max_extents(len);
7844
7845         /*
7846          * If we have an outstanding_extents count still set then we're
7847          * within our reservation, otherwise we need to adjust our inode
7848          * counter appropriately.
7849          */
7850         if (dio_data->outstanding_extents >= num_extents) {
7851                 dio_data->outstanding_extents -= num_extents;
7852         } else {
7853                 /*
7854                  * If dio write length has been split due to no large enough
7855                  * contiguous space, we need to compensate our inode counter
7856                  * appropriately.
7857                  */
7858                 u64 num_needed = num_extents - dio_data->outstanding_extents;
7859
7860                 spin_lock(&BTRFS_I(inode)->lock);
7861                 BTRFS_I(inode)->outstanding_extents += num_needed;
7862                 spin_unlock(&BTRFS_I(inode)->lock);
7863         }
7864 }
7865
7866 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7867                                    struct buffer_head *bh_result, int create)
7868 {
7869         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7870         struct extent_map *em;
7871         struct extent_state *cached_state = NULL;
7872         struct btrfs_dio_data *dio_data = NULL;
7873         u64 start = iblock << inode->i_blkbits;
7874         u64 lockstart, lockend;
7875         u64 len = bh_result->b_size;
7876         int unlock_bits = EXTENT_LOCKED;
7877         int ret = 0;
7878
7879         if (create)
7880                 unlock_bits |= EXTENT_DIRTY;
7881         else
7882                 len = min_t(u64, len, fs_info->sectorsize);
7883
7884         lockstart = start;
7885         lockend = start + len - 1;
7886
7887         if (current->journal_info) {
7888                 /*
7889                  * Need to pull our outstanding extents and set journal_info to NULL so
7890                  * that anything that needs to check if there's a transaction doesn't get
7891                  * confused.
7892                  */
7893                 dio_data = current->journal_info;
7894                 current->journal_info = NULL;
7895         }
7896
7897         /*
7898          * If this errors out it's because we couldn't invalidate pagecache for
7899          * this range and we need to fallback to buffered.
7900          */
7901         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7902                                create)) {
7903                 ret = -ENOTBLK;
7904                 goto err;
7905         }
7906
7907         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7908         if (IS_ERR(em)) {
7909                 ret = PTR_ERR(em);
7910                 goto unlock_err;
7911         }
7912
7913         /*
7914          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7915          * io.  INLINE is special, and we could probably kludge it in here, but
7916          * it's still buffered so for safety lets just fall back to the generic
7917          * buffered path.
7918          *
7919          * For COMPRESSED we _have_ to read the entire extent in so we can
7920          * decompress it, so there will be buffering required no matter what we
7921          * do, so go ahead and fallback to buffered.
7922          *
7923          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7924          * to buffered IO.  Don't blame me, this is the price we pay for using
7925          * the generic code.
7926          */
7927         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7928             em->block_start == EXTENT_MAP_INLINE) {
7929                 free_extent_map(em);
7930                 ret = -ENOTBLK;
7931                 goto unlock_err;
7932         }
7933
7934         /* Just a good old fashioned hole, return */
7935         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7936                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7937                 free_extent_map(em);
7938                 goto unlock_err;
7939         }
7940
7941         /*
7942          * We don't allocate a new extent in the following cases
7943          *
7944          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7945          * existing extent.
7946          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7947          * just use the extent.
7948          *
7949          */
7950         if (!create) {
7951                 len = min(len, em->len - (start - em->start));
7952                 lockstart = start + len;
7953                 goto unlock;
7954         }
7955
7956         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7957             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7958              em->block_start != EXTENT_MAP_HOLE)) {
7959                 int type;
7960                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7961
7962                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7963                         type = BTRFS_ORDERED_PREALLOC;
7964                 else
7965                         type = BTRFS_ORDERED_NOCOW;
7966                 len = min(len, em->len - (start - em->start));
7967                 block_start = em->block_start + (start - em->start);
7968
7969                 if (can_nocow_extent(inode, start, &len, &orig_start,
7970                                      &orig_block_len, &ram_bytes) == 1 &&
7971                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7972                         struct extent_map *em2;
7973
7974                         em2 = btrfs_create_dio_extent(inode, start, len,
7975                                                       orig_start, block_start,
7976                                                       len, orig_block_len,
7977                                                       ram_bytes, type);
7978                         btrfs_dec_nocow_writers(fs_info, block_start);
7979                         if (type == BTRFS_ORDERED_PREALLOC) {
7980                                 free_extent_map(em);
7981                                 em = em2;
7982                         }
7983                         if (em2 && IS_ERR(em2)) {
7984                                 ret = PTR_ERR(em2);
7985                                 goto unlock_err;
7986                         }
7987                         /*
7988                          * For inode marked NODATACOW or extent marked PREALLOC,
7989                          * use the existing or preallocated extent, so does not
7990                          * need to adjust btrfs_space_info's bytes_may_use.
7991                          */
7992                         btrfs_free_reserved_data_space_noquota(inode,
7993                                         start, len);
7994                         goto unlock;
7995                 }
7996         }
7997
7998         /*
7999          * this will cow the extent, reset the len in case we changed
8000          * it above
8001          */
8002         len = bh_result->b_size;
8003         free_extent_map(em);
8004         em = btrfs_new_extent_direct(inode, start, len);
8005         if (IS_ERR(em)) {
8006                 ret = PTR_ERR(em);
8007                 goto unlock_err;
8008         }
8009         len = min(len, em->len - (start - em->start));
8010 unlock:
8011         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
8012                 inode->i_blkbits;
8013         bh_result->b_size = len;
8014         bh_result->b_bdev = em->bdev;
8015         set_buffer_mapped(bh_result);
8016         if (create) {
8017                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
8018                         set_buffer_new(bh_result);
8019
8020                 /*
8021                  * Need to update the i_size under the extent lock so buffered
8022                  * readers will get the updated i_size when we unlock.
8023                  */
8024                 if (!dio_data->overwrite && start + len > i_size_read(inode))
8025                         i_size_write(inode, start + len);
8026
8027                 adjust_dio_outstanding_extents(inode, dio_data, len);
8028                 WARN_ON(dio_data->reserve < len);
8029                 dio_data->reserve -= len;
8030                 dio_data->unsubmitted_oe_range_end = start + len;
8031                 current->journal_info = dio_data;
8032         }
8033
8034         /*
8035          * In the case of write we need to clear and unlock the entire range,
8036          * in the case of read we need to unlock only the end area that we
8037          * aren't using if there is any left over space.
8038          */
8039         if (lockstart < lockend) {
8040                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
8041                                  lockend, unlock_bits, 1, 0,
8042                                  &cached_state, GFP_NOFS);
8043         } else {
8044                 free_extent_state(cached_state);
8045         }
8046
8047         free_extent_map(em);
8048
8049         return 0;
8050
8051 unlock_err:
8052         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
8053                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
8054 err:
8055         if (dio_data)
8056                 current->journal_info = dio_data;
8057         /*
8058          * Compensate the delalloc release we do in btrfs_direct_IO() when we
8059          * write less data then expected, so that we don't underflow our inode's
8060          * outstanding extents counter.
8061          */
8062         if (create && dio_data)
8063                 adjust_dio_outstanding_extents(inode, dio_data, len);
8064
8065         return ret;
8066 }
8067
8068 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
8069                                                  struct bio *bio,
8070                                                  int mirror_num)
8071 {
8072         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8073         blk_status_t ret;
8074
8075         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
8076
8077         bio_get(bio);
8078
8079         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
8080         if (ret)
8081                 goto err;
8082
8083         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
8084 err:
8085         bio_put(bio);
8086         return ret;
8087 }
8088
8089 static int btrfs_check_dio_repairable(struct inode *inode,
8090                                       struct bio *failed_bio,
8091                                       struct io_failure_record *failrec,
8092                                       int failed_mirror)
8093 {
8094         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8095         int num_copies;
8096
8097         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
8098         if (num_copies == 1) {
8099                 /*
8100                  * we only have a single copy of the data, so don't bother with
8101                  * all the retry and error correction code that follows. no
8102                  * matter what the error is, it is very likely to persist.
8103                  */
8104                 btrfs_debug(fs_info,
8105                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
8106                         num_copies, failrec->this_mirror, failed_mirror);
8107                 return 0;
8108         }
8109
8110         failrec->failed_mirror = failed_mirror;
8111         failrec->this_mirror++;
8112         if (failrec->this_mirror == failed_mirror)
8113                 failrec->this_mirror++;
8114
8115         if (failrec->this_mirror > num_copies) {
8116                 btrfs_debug(fs_info,
8117                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
8118                         num_copies, failrec->this_mirror, failed_mirror);
8119                 return 0;
8120         }
8121
8122         return 1;
8123 }
8124
8125 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
8126                                    struct page *page, unsigned int pgoff,
8127                                    u64 start, u64 end, int failed_mirror,
8128                                    bio_end_io_t *repair_endio, void *repair_arg)
8129 {
8130         struct io_failure_record *failrec;
8131         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8132         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
8133         struct bio *bio;
8134         int isector;
8135         unsigned int read_mode = 0;
8136         int segs;
8137         int ret;
8138         blk_status_t status;
8139
8140         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
8141
8142         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
8143         if (ret)
8144                 return errno_to_blk_status(ret);
8145
8146         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
8147                                          failed_mirror);
8148         if (!ret) {
8149                 free_io_failure(failure_tree, io_tree, failrec);
8150                 return BLK_STS_IOERR;
8151         }
8152
8153         segs = bio_segments(failed_bio);
8154         if (segs > 1 ||
8155             (failed_bio->bi_io_vec->bv_len > btrfs_inode_sectorsize(inode)))
8156                 read_mode |= REQ_FAILFAST_DEV;
8157
8158         isector = start - btrfs_io_bio(failed_bio)->logical;
8159         isector >>= inode->i_sb->s_blocksize_bits;
8160         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
8161                                 pgoff, isector, repair_endio, repair_arg);
8162         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
8163
8164         btrfs_debug(BTRFS_I(inode)->root->fs_info,
8165                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
8166                     read_mode, failrec->this_mirror, failrec->in_validation);
8167
8168         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
8169         if (status) {
8170                 free_io_failure(failure_tree, io_tree, failrec);
8171                 bio_put(bio);
8172         }
8173
8174         return status;
8175 }
8176
8177 struct btrfs_retry_complete {
8178         struct completion done;
8179         struct inode *inode;
8180         u64 start;
8181         int uptodate;
8182 };
8183
8184 static void btrfs_retry_endio_nocsum(struct bio *bio)
8185 {
8186         struct btrfs_retry_complete *done = bio->bi_private;
8187         struct inode *inode = done->inode;
8188         struct bio_vec *bvec;
8189         struct extent_io_tree *io_tree, *failure_tree;
8190         int i;
8191
8192         if (bio->bi_status)
8193                 goto end;
8194
8195         ASSERT(bio->bi_vcnt == 1);
8196         io_tree = &BTRFS_I(inode)->io_tree;
8197         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8198         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(inode));
8199
8200         done->uptodate = 1;
8201         ASSERT(!bio_flagged(bio, BIO_CLONED));
8202         bio_for_each_segment_all(bvec, bio, i)
8203                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
8204                                  io_tree, done->start, bvec->bv_page,
8205                                  btrfs_ino(BTRFS_I(inode)), 0);
8206 end:
8207         complete(&done->done);
8208         bio_put(bio);
8209 }
8210
8211 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
8212                                                 struct btrfs_io_bio *io_bio)
8213 {
8214         struct btrfs_fs_info *fs_info;
8215         struct bio_vec bvec;
8216         struct bvec_iter iter;
8217         struct btrfs_retry_complete done;
8218         u64 start;
8219         unsigned int pgoff;
8220         u32 sectorsize;
8221         int nr_sectors;
8222         blk_status_t ret;
8223         blk_status_t err = BLK_STS_OK;
8224
8225         fs_info = BTRFS_I(inode)->root->fs_info;
8226         sectorsize = fs_info->sectorsize;
8227
8228         start = io_bio->logical;
8229         done.inode = inode;
8230         io_bio->bio.bi_iter = io_bio->iter;
8231
8232         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8233                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8234                 pgoff = bvec.bv_offset;
8235
8236 next_block_or_try_again:
8237                 done.uptodate = 0;
8238                 done.start = start;
8239                 init_completion(&done.done);
8240
8241                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8242                                 pgoff, start, start + sectorsize - 1,
8243                                 io_bio->mirror_num,
8244                                 btrfs_retry_endio_nocsum, &done);
8245                 if (ret) {
8246                         err = ret;
8247                         goto next;
8248                 }
8249
8250                 wait_for_completion_io(&done.done);
8251
8252                 if (!done.uptodate) {
8253                         /* We might have another mirror, so try again */
8254                         goto next_block_or_try_again;
8255                 }
8256
8257 next:
8258                 start += sectorsize;
8259
8260                 nr_sectors--;
8261                 if (nr_sectors) {
8262                         pgoff += sectorsize;
8263                         ASSERT(pgoff < PAGE_SIZE);
8264                         goto next_block_or_try_again;
8265                 }
8266         }
8267
8268         return err;
8269 }
8270
8271 static void btrfs_retry_endio(struct bio *bio)
8272 {
8273         struct btrfs_retry_complete *done = bio->bi_private;
8274         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8275         struct extent_io_tree *io_tree, *failure_tree;
8276         struct inode *inode = done->inode;
8277         struct bio_vec *bvec;
8278         int uptodate;
8279         int ret;
8280         int i;
8281
8282         if (bio->bi_status)
8283                 goto end;
8284
8285         uptodate = 1;
8286
8287         ASSERT(bio->bi_vcnt == 1);
8288         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(done->inode));
8289
8290         io_tree = &BTRFS_I(inode)->io_tree;
8291         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8292
8293         ASSERT(!bio_flagged(bio, BIO_CLONED));
8294         bio_for_each_segment_all(bvec, bio, i) {
8295                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
8296                                              bvec->bv_offset, done->start,
8297                                              bvec->bv_len);
8298                 if (!ret)
8299                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
8300                                          failure_tree, io_tree, done->start,
8301                                          bvec->bv_page,
8302                                          btrfs_ino(BTRFS_I(inode)),
8303                                          bvec->bv_offset);
8304                 else
8305                         uptodate = 0;
8306         }
8307
8308         done->uptodate = uptodate;
8309 end:
8310         complete(&done->done);
8311         bio_put(bio);
8312 }
8313
8314 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
8315                 struct btrfs_io_bio *io_bio, blk_status_t err)
8316 {
8317         struct btrfs_fs_info *fs_info;
8318         struct bio_vec bvec;
8319         struct bvec_iter iter;
8320         struct btrfs_retry_complete done;
8321         u64 start;
8322         u64 offset = 0;
8323         u32 sectorsize;
8324         int nr_sectors;
8325         unsigned int pgoff;
8326         int csum_pos;
8327         bool uptodate = (err == 0);
8328         int ret;
8329         blk_status_t status;
8330
8331         fs_info = BTRFS_I(inode)->root->fs_info;
8332         sectorsize = fs_info->sectorsize;
8333
8334         err = BLK_STS_OK;
8335         start = io_bio->logical;
8336         done.inode = inode;
8337         io_bio->bio.bi_iter = io_bio->iter;
8338
8339         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8340                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8341
8342                 pgoff = bvec.bv_offset;
8343 next_block:
8344                 if (uptodate) {
8345                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8346                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
8347                                         bvec.bv_page, pgoff, start, sectorsize);
8348                         if (likely(!ret))
8349                                 goto next;
8350                 }
8351 try_again:
8352                 done.uptodate = 0;
8353                 done.start = start;
8354                 init_completion(&done.done);
8355
8356                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8357                                         pgoff, start, start + sectorsize - 1,
8358                                         io_bio->mirror_num, btrfs_retry_endio,
8359                                         &done);
8360                 if (status) {
8361                         err = status;
8362                         goto next;
8363                 }
8364
8365                 wait_for_completion_io(&done.done);
8366
8367                 if (!done.uptodate) {
8368                         /* We might have another mirror, so try again */
8369                         goto try_again;
8370                 }
8371 next:
8372                 offset += sectorsize;
8373                 start += sectorsize;
8374
8375                 ASSERT(nr_sectors);
8376
8377                 nr_sectors--;
8378                 if (nr_sectors) {
8379                         pgoff += sectorsize;
8380                         ASSERT(pgoff < PAGE_SIZE);
8381                         goto next_block;
8382                 }
8383         }
8384
8385         return err;
8386 }
8387
8388 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8389                 struct btrfs_io_bio *io_bio, blk_status_t err)
8390 {
8391         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8392
8393         if (skip_csum) {
8394                 if (unlikely(err))
8395                         return __btrfs_correct_data_nocsum(inode, io_bio);
8396                 else
8397                         return BLK_STS_OK;
8398         } else {
8399                 return __btrfs_subio_endio_read(inode, io_bio, err);
8400         }
8401 }
8402
8403 static void btrfs_endio_direct_read(struct bio *bio)
8404 {
8405         struct btrfs_dio_private *dip = bio->bi_private;
8406         struct inode *inode = dip->inode;
8407         struct bio *dio_bio;
8408         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8409         blk_status_t err = bio->bi_status;
8410
8411         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8412                 err = btrfs_subio_endio_read(inode, io_bio, err);
8413
8414         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8415                       dip->logical_offset + dip->bytes - 1);
8416         dio_bio = dip->dio_bio;
8417
8418         kfree(dip);
8419
8420         dio_bio->bi_status = err;
8421         dio_end_io(dio_bio);
8422
8423         if (io_bio->end_io)
8424                 io_bio->end_io(io_bio, blk_status_to_errno(err));
8425         bio_put(bio);
8426 }
8427
8428 static void __endio_write_update_ordered(struct inode *inode,
8429                                          const u64 offset, const u64 bytes,
8430                                          const bool uptodate)
8431 {
8432         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8433         struct btrfs_ordered_extent *ordered = NULL;
8434         struct btrfs_workqueue *wq;
8435         btrfs_work_func_t func;
8436         u64 ordered_offset = offset;
8437         u64 ordered_bytes = bytes;
8438         u64 last_offset;
8439         int ret;
8440
8441         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
8442                 wq = fs_info->endio_freespace_worker;
8443                 func = btrfs_freespace_write_helper;
8444         } else {
8445                 wq = fs_info->endio_write_workers;
8446                 func = btrfs_endio_write_helper;
8447         }
8448
8449 again:
8450         last_offset = ordered_offset;
8451         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8452                                                    &ordered_offset,
8453                                                    ordered_bytes,
8454                                                    uptodate);
8455         if (!ret)
8456                 goto out_test;
8457
8458         btrfs_init_work(&ordered->work, func, finish_ordered_fn, NULL, NULL);
8459         btrfs_queue_work(wq, &ordered->work);
8460 out_test:
8461         /*
8462          * If btrfs_dec_test_ordered_pending does not find any ordered extent
8463          * in the range, we can exit.
8464          */
8465         if (ordered_offset == last_offset)
8466                 return;
8467         /*
8468          * our bio might span multiple ordered extents.  If we haven't
8469          * completed the accounting for the whole dio, go back and try again
8470          */
8471         if (ordered_offset < offset + bytes) {
8472                 ordered_bytes = offset + bytes - ordered_offset;
8473                 ordered = NULL;
8474                 goto again;
8475         }
8476 }
8477
8478 static void btrfs_endio_direct_write(struct bio *bio)
8479 {
8480         struct btrfs_dio_private *dip = bio->bi_private;
8481         struct bio *dio_bio = dip->dio_bio;
8482
8483         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8484                                      dip->bytes, !bio->bi_status);
8485
8486         kfree(dip);
8487
8488         dio_bio->bi_status = bio->bi_status;
8489         dio_end_io(dio_bio);
8490         bio_put(bio);
8491 }
8492
8493 static blk_status_t __btrfs_submit_bio_start_direct_io(void *private_data,
8494                                     struct bio *bio, int mirror_num,
8495                                     unsigned long bio_flags, u64 offset)
8496 {
8497         struct inode *inode = private_data;
8498         blk_status_t ret;
8499         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8500         BUG_ON(ret); /* -ENOMEM */
8501         return 0;
8502 }
8503
8504 static void btrfs_end_dio_bio(struct bio *bio)
8505 {
8506         struct btrfs_dio_private *dip = bio->bi_private;
8507         blk_status_t err = bio->bi_status;
8508
8509         if (err)
8510                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8511                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8512                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8513                            bio->bi_opf,
8514                            (unsigned long long)bio->bi_iter.bi_sector,
8515                            bio->bi_iter.bi_size, err);
8516
8517         if (dip->subio_endio)
8518                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8519
8520         if (err) {
8521                 dip->errors = 1;
8522
8523                 /*
8524                  * before atomic variable goto zero, we must make sure
8525                  * dip->errors is perceived to be set.
8526                  */
8527                 smp_mb__before_atomic();
8528         }
8529
8530         /* if there are more bios still pending for this dio, just exit */
8531         if (!atomic_dec_and_test(&dip->pending_bios))
8532                 goto out;
8533
8534         if (dip->errors) {
8535                 bio_io_error(dip->orig_bio);
8536         } else {
8537                 dip->dio_bio->bi_status = 0;
8538                 bio_endio(dip->orig_bio);
8539         }
8540 out:
8541         bio_put(bio);
8542 }
8543
8544 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8545                                                  struct btrfs_dio_private *dip,
8546                                                  struct bio *bio,
8547                                                  u64 file_offset)
8548 {
8549         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8550         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8551         blk_status_t ret;
8552
8553         /*
8554          * We load all the csum data we need when we submit
8555          * the first bio to reduce the csum tree search and
8556          * contention.
8557          */
8558         if (dip->logical_offset == file_offset) {
8559                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8560                                                 file_offset);
8561                 if (ret)
8562                         return ret;
8563         }
8564
8565         if (bio == dip->orig_bio)
8566                 return 0;
8567
8568         file_offset -= dip->logical_offset;
8569         file_offset >>= inode->i_sb->s_blocksize_bits;
8570         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8571
8572         return 0;
8573 }
8574
8575 static inline blk_status_t
8576 __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode, u64 file_offset,
8577                        int async_submit)
8578 {
8579         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8580         struct btrfs_dio_private *dip = bio->bi_private;
8581         bool write = bio_op(bio) == REQ_OP_WRITE;
8582         blk_status_t ret;
8583
8584         if (async_submit)
8585                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8586
8587         bio_get(bio);
8588
8589         if (!write) {
8590                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8591                 if (ret)
8592                         goto err;
8593         }
8594
8595         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8596                 goto map;
8597
8598         if (write && async_submit) {
8599                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8600                                           file_offset, inode,
8601                                           __btrfs_submit_bio_start_direct_io,
8602                                           __btrfs_submit_bio_done);
8603                 goto err;
8604         } else if (write) {
8605                 /*
8606                  * If we aren't doing async submit, calculate the csum of the
8607                  * bio now.
8608                  */
8609                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8610                 if (ret)
8611                         goto err;
8612         } else {
8613                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8614                                                      file_offset);
8615                 if (ret)
8616                         goto err;
8617         }
8618 map:
8619         ret = btrfs_map_bio(fs_info, bio, 0, async_submit);
8620 err:
8621         bio_put(bio);
8622         return ret;
8623 }
8624
8625 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8626 {
8627         struct inode *inode = dip->inode;
8628         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8629         struct bio *bio;
8630         struct bio *orig_bio = dip->orig_bio;
8631         u64 start_sector = orig_bio->bi_iter.bi_sector;
8632         u64 file_offset = dip->logical_offset;
8633         u64 map_length;
8634         int async_submit = 0;
8635         u64 submit_len;
8636         int clone_offset = 0;
8637         int clone_len;
8638         int ret;
8639         blk_status_t status;
8640
8641         map_length = orig_bio->bi_iter.bi_size;
8642         submit_len = map_length;
8643         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio), start_sector << 9,
8644                               &map_length, NULL, 0);
8645         if (ret)
8646                 return -EIO;
8647
8648         if (map_length >= submit_len) {
8649                 bio = orig_bio;
8650                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8651                 goto submit;
8652         }
8653
8654         /* async crcs make it difficult to collect full stripe writes. */
8655         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8656                 async_submit = 0;
8657         else
8658                 async_submit = 1;
8659
8660         /* bio split */
8661         ASSERT(map_length <= INT_MAX);
8662         atomic_inc(&dip->pending_bios);
8663         do {
8664                 clone_len = min_t(int, submit_len, map_length);
8665
8666                 /*
8667                  * This will never fail as it's passing GPF_NOFS and
8668                  * the allocation is backed by btrfs_bioset.
8669                  */
8670                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8671                                               clone_len);
8672                 bio->bi_private = dip;
8673                 bio->bi_end_io = btrfs_end_dio_bio;
8674                 btrfs_io_bio(bio)->logical = file_offset;
8675
8676                 ASSERT(submit_len >= clone_len);
8677                 submit_len -= clone_len;
8678                 if (submit_len == 0)
8679                         break;
8680
8681                 /*
8682                  * Increase the count before we submit the bio so we know
8683                  * the end IO handler won't happen before we increase the
8684                  * count. Otherwise, the dip might get freed before we're
8685                  * done setting it up.
8686                  */
8687                 atomic_inc(&dip->pending_bios);
8688
8689                 status = __btrfs_submit_dio_bio(bio, inode, file_offset,
8690                                                 async_submit);
8691                 if (status) {
8692                         bio_put(bio);
8693                         atomic_dec(&dip->pending_bios);
8694                         goto out_err;
8695                 }
8696
8697                 clone_offset += clone_len;
8698                 start_sector += clone_len >> 9;
8699                 file_offset += clone_len;
8700
8701                 map_length = submit_len;
8702                 ret = btrfs_map_block(fs_info, btrfs_op(orig_bio),
8703                                       start_sector << 9, &map_length, NULL, 0);
8704                 if (ret)
8705                         goto out_err;
8706         } while (submit_len > 0);
8707
8708 submit:
8709         status = __btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8710         if (!status)
8711                 return 0;
8712
8713         bio_put(bio);
8714 out_err:
8715         dip->errors = 1;
8716         /*
8717          * before atomic variable goto zero, we must
8718          * make sure dip->errors is perceived to be set.
8719          */
8720         smp_mb__before_atomic();
8721         if (atomic_dec_and_test(&dip->pending_bios))
8722                 bio_io_error(dip->orig_bio);
8723
8724         /* bio_end_io() will handle error, so we needn't return it */
8725         return 0;
8726 }
8727
8728 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8729                                 loff_t file_offset)
8730 {
8731         struct btrfs_dio_private *dip = NULL;
8732         struct bio *bio = NULL;
8733         struct btrfs_io_bio *io_bio;
8734         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8735         int ret = 0;
8736
8737         bio = btrfs_bio_clone(dio_bio);
8738
8739         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8740         if (!dip) {
8741                 ret = -ENOMEM;
8742                 goto free_ordered;
8743         }
8744
8745         dip->private = dio_bio->bi_private;
8746         dip->inode = inode;
8747         dip->logical_offset = file_offset;
8748         dip->bytes = dio_bio->bi_iter.bi_size;
8749         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8750         bio->bi_private = dip;
8751         dip->orig_bio = bio;
8752         dip->dio_bio = dio_bio;
8753         atomic_set(&dip->pending_bios, 0);
8754         io_bio = btrfs_io_bio(bio);
8755         io_bio->logical = file_offset;
8756
8757         if (write) {
8758                 bio->bi_end_io = btrfs_endio_direct_write;
8759         } else {
8760                 bio->bi_end_io = btrfs_endio_direct_read;
8761                 dip->subio_endio = btrfs_subio_endio_read;
8762         }
8763
8764         /*
8765          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8766          * even if we fail to submit a bio, because in such case we do the
8767          * corresponding error handling below and it must not be done a second
8768          * time by btrfs_direct_IO().
8769          */
8770         if (write) {
8771                 struct btrfs_dio_data *dio_data = current->journal_info;
8772
8773                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8774                         dip->bytes;
8775                 dio_data->unsubmitted_oe_range_start =
8776                         dio_data->unsubmitted_oe_range_end;
8777         }
8778
8779         ret = btrfs_submit_direct_hook(dip);
8780         if (!ret)
8781                 return;
8782
8783         if (io_bio->end_io)
8784                 io_bio->end_io(io_bio, ret);
8785
8786 free_ordered:
8787         /*
8788          * If we arrived here it means either we failed to submit the dip
8789          * or we either failed to clone the dio_bio or failed to allocate the
8790          * dip. If we cloned the dio_bio and allocated the dip, we can just
8791          * call bio_endio against our io_bio so that we get proper resource
8792          * cleanup if we fail to submit the dip, otherwise, we must do the
8793          * same as btrfs_endio_direct_[write|read] because we can't call these
8794          * callbacks - they require an allocated dip and a clone of dio_bio.
8795          */
8796         if (bio && dip) {
8797                 bio_io_error(bio);
8798                 /*
8799                  * The end io callbacks free our dip, do the final put on bio
8800                  * and all the cleanup and final put for dio_bio (through
8801                  * dio_end_io()).
8802                  */
8803                 dip = NULL;
8804                 bio = NULL;
8805         } else {
8806                 if (write)
8807                         __endio_write_update_ordered(inode,
8808                                                 file_offset,
8809                                                 dio_bio->bi_iter.bi_size,
8810                                                 false);
8811                 else
8812                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8813                               file_offset + dio_bio->bi_iter.bi_size - 1);
8814
8815                 dio_bio->bi_status = BLK_STS_IOERR;
8816                 /*
8817                  * Releases and cleans up our dio_bio, no need to bio_put()
8818                  * nor bio_endio()/bio_io_error() against dio_bio.
8819                  */
8820                 dio_end_io(dio_bio);
8821         }
8822         if (bio)
8823                 bio_put(bio);
8824         kfree(dip);
8825 }
8826
8827 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8828                                struct kiocb *iocb,
8829                                const struct iov_iter *iter, loff_t offset)
8830 {
8831         int seg;
8832         int i;
8833         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8834         ssize_t retval = -EINVAL;
8835
8836         if (offset & blocksize_mask)
8837                 goto out;
8838
8839         if (iov_iter_alignment(iter) & blocksize_mask)
8840                 goto out;
8841
8842         /* If this is a write we don't need to check anymore */
8843         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8844                 return 0;
8845         /*
8846          * Check to make sure we don't have duplicate iov_base's in this
8847          * iovec, if so return EINVAL, otherwise we'll get csum errors
8848          * when reading back.
8849          */
8850         for (seg = 0; seg < iter->nr_segs; seg++) {
8851                 for (i = seg + 1; i < iter->nr_segs; i++) {
8852                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8853                                 goto out;
8854                 }
8855         }
8856         retval = 0;
8857 out:
8858         return retval;
8859 }
8860
8861 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8862 {
8863         struct file *file = iocb->ki_filp;
8864         struct inode *inode = file->f_mapping->host;
8865         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8866         struct btrfs_dio_data dio_data = { 0 };
8867         struct extent_changeset *data_reserved = NULL;
8868         loff_t offset = iocb->ki_pos;
8869         size_t count = 0;
8870         int flags = 0;
8871         bool wakeup = true;
8872         bool relock = false;
8873         ssize_t ret;
8874
8875         if (check_direct_IO(fs_info, iocb, iter, offset))
8876                 return 0;
8877
8878         inode_dio_begin(inode);
8879
8880         /*
8881          * The generic stuff only does filemap_write_and_wait_range, which
8882          * isn't enough if we've written compressed pages to this area, so
8883          * we need to flush the dirty pages again to make absolutely sure
8884          * that any outstanding dirty pages are on disk.
8885          */
8886         count = iov_iter_count(iter);
8887         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8888                      &BTRFS_I(inode)->runtime_flags))
8889                 filemap_fdatawrite_range(inode->i_mapping, offset,
8890                                          offset + count - 1);
8891
8892         if (iov_iter_rw(iter) == WRITE) {
8893                 /*
8894                  * If the write DIO is beyond the EOF, we need update
8895                  * the isize, but it is protected by i_mutex. So we can
8896                  * not unlock the i_mutex at this case.
8897                  */
8898                 if (offset + count <= inode->i_size) {
8899                         dio_data.overwrite = 1;
8900                         inode_unlock(inode);
8901                         relock = true;
8902                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8903                         ret = -EAGAIN;
8904                         goto out;
8905                 }
8906                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8907                                                    offset, count);
8908                 if (ret)
8909                         goto out;
8910                 dio_data.outstanding_extents = count_max_extents(count);
8911
8912                 /*
8913                  * We need to know how many extents we reserved so that we can
8914                  * do the accounting properly if we go over the number we
8915                  * originally calculated.  Abuse current->journal_info for this.
8916                  */
8917                 dio_data.reserve = round_up(count,
8918                                             fs_info->sectorsize);
8919                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8920                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8921                 current->journal_info = &dio_data;
8922                 down_read(&BTRFS_I(inode)->dio_sem);
8923         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8924                                      &BTRFS_I(inode)->runtime_flags)) {
8925                 inode_dio_end(inode);
8926                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8927                 wakeup = false;
8928         }
8929
8930         ret = __blockdev_direct_IO(iocb, inode,
8931                                    fs_info->fs_devices->latest_bdev,
8932                                    iter, btrfs_get_blocks_direct, NULL,
8933                                    btrfs_submit_direct, flags);
8934         if (iov_iter_rw(iter) == WRITE) {
8935                 up_read(&BTRFS_I(inode)->dio_sem);
8936                 current->journal_info = NULL;
8937                 if (ret < 0 && ret != -EIOCBQUEUED) {
8938                         if (dio_data.reserve)
8939                                 btrfs_delalloc_release_space(inode, data_reserved,
8940                                         offset, dio_data.reserve);
8941                         /*
8942                          * On error we might have left some ordered extents
8943                          * without submitting corresponding bios for them, so
8944                          * cleanup them up to avoid other tasks getting them
8945                          * and waiting for them to complete forever.
8946                          */
8947                         if (dio_data.unsubmitted_oe_range_start <
8948                             dio_data.unsubmitted_oe_range_end)
8949                                 __endio_write_update_ordered(inode,
8950                                         dio_data.unsubmitted_oe_range_start,
8951                                         dio_data.unsubmitted_oe_range_end -
8952                                         dio_data.unsubmitted_oe_range_start,
8953                                         false);
8954                 } else if (ret >= 0 && (size_t)ret < count)
8955                         btrfs_delalloc_release_space(inode, data_reserved,
8956                                         offset, count - (size_t)ret);
8957         }
8958 out:
8959         if (wakeup)
8960                 inode_dio_end(inode);
8961         if (relock)
8962                 inode_lock(inode);
8963
8964         extent_changeset_free(data_reserved);
8965         return ret;
8966 }
8967
8968 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8969
8970 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8971                 __u64 start, __u64 len)
8972 {
8973         int     ret;
8974
8975         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8976         if (ret)
8977                 return ret;
8978
8979         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8980 }
8981
8982 int btrfs_readpage(struct file *file, struct page *page)
8983 {
8984         struct extent_io_tree *tree;
8985         tree = &BTRFS_I(page->mapping->host)->io_tree;
8986         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8987 }
8988
8989 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8990 {
8991         struct extent_io_tree *tree;
8992         struct inode *inode = page->mapping->host;
8993         int ret;
8994
8995         if (current->flags & PF_MEMALLOC) {
8996                 redirty_page_for_writepage(wbc, page);
8997                 unlock_page(page);
8998                 return 0;
8999         }
9000
9001         /*
9002          * If we are under memory pressure we will call this directly from the
9003          * VM, we need to make sure we have the inode referenced for the ordered
9004          * extent.  If not just return like we didn't do anything.
9005          */
9006         if (!igrab(inode)) {
9007                 redirty_page_for_writepage(wbc, page);
9008                 return AOP_WRITEPAGE_ACTIVATE;
9009         }
9010         tree = &BTRFS_I(page->mapping->host)->io_tree;
9011         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
9012         btrfs_add_delayed_iput(inode);
9013         return ret;
9014 }
9015
9016 static int btrfs_writepages(struct address_space *mapping,
9017                             struct writeback_control *wbc)
9018 {
9019         struct extent_io_tree *tree;
9020
9021         tree = &BTRFS_I(mapping->host)->io_tree;
9022         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
9023 }
9024
9025 static int
9026 btrfs_readpages(struct file *file, struct address_space *mapping,
9027                 struct list_head *pages, unsigned nr_pages)
9028 {
9029         struct extent_io_tree *tree;
9030         tree = &BTRFS_I(mapping->host)->io_tree;
9031         return extent_readpages(tree, mapping, pages, nr_pages,
9032                                 btrfs_get_extent);
9033 }
9034 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
9035 {
9036         struct extent_io_tree *tree;
9037         struct extent_map_tree *map;
9038         int ret;
9039
9040         tree = &BTRFS_I(page->mapping->host)->io_tree;
9041         map = &BTRFS_I(page->mapping->host)->extent_tree;
9042         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
9043         if (ret == 1) {
9044                 ClearPagePrivate(page);
9045                 set_page_private(page, 0);
9046                 put_page(page);
9047         }
9048         return ret;
9049 }
9050
9051 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
9052 {
9053         if (PageWriteback(page) || PageDirty(page))
9054                 return 0;
9055         return __btrfs_releasepage(page, gfp_flags);
9056 }
9057
9058 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
9059                                  unsigned int length)
9060 {
9061         struct inode *inode = page->mapping->host;
9062         struct extent_io_tree *tree;
9063         struct btrfs_ordered_extent *ordered;
9064         struct extent_state *cached_state = NULL;
9065         u64 page_start = page_offset(page);
9066         u64 page_end = page_start + PAGE_SIZE - 1;
9067         u64 start;
9068         u64 end;
9069         int inode_evicting = inode->i_state & I_FREEING;
9070
9071         /*
9072          * we have the page locked, so new writeback can't start,
9073          * and the dirty bit won't be cleared while we are here.
9074          *
9075          * Wait for IO on this page so that we can safely clear
9076          * the PagePrivate2 bit and do ordered accounting
9077          */
9078         wait_on_page_writeback(page);
9079
9080         tree = &BTRFS_I(inode)->io_tree;
9081         if (offset) {
9082                 btrfs_releasepage(page, GFP_NOFS);
9083                 return;
9084         }
9085
9086         if (!inode_evicting)
9087                 lock_extent_bits(tree, page_start, page_end, &cached_state);
9088 again:
9089         start = page_start;
9090         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
9091                                         page_end - start + 1);
9092         if (ordered) {
9093                 end = min(page_end, ordered->file_offset + ordered->len - 1);
9094                 /*
9095                  * IO on this page will never be started, so we need
9096                  * to account for any ordered extents now
9097                  */
9098                 if (!inode_evicting)
9099                         clear_extent_bit(tree, start, end,
9100                                          EXTENT_DIRTY | EXTENT_DELALLOC |
9101                                          EXTENT_DELALLOC_NEW |
9102                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
9103                                          EXTENT_DEFRAG, 1, 0, &cached_state,
9104                                          GFP_NOFS);
9105                 /*
9106                  * whoever cleared the private bit is responsible
9107                  * for the finish_ordered_io
9108                  */
9109                 if (TestClearPagePrivate2(page)) {
9110                         struct btrfs_ordered_inode_tree *tree;
9111                         u64 new_len;
9112
9113                         tree = &BTRFS_I(inode)->ordered_tree;
9114
9115                         spin_lock_irq(&tree->lock);
9116                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
9117                         new_len = start - ordered->file_offset;
9118                         if (new_len < ordered->truncated_len)
9119                                 ordered->truncated_len = new_len;
9120                         spin_unlock_irq(&tree->lock);
9121
9122                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
9123                                                            start,
9124                                                            end - start + 1, 1))
9125                                 btrfs_finish_ordered_io(ordered);
9126                 }
9127                 btrfs_put_ordered_extent(ordered);
9128                 if (!inode_evicting) {
9129                         cached_state = NULL;
9130                         lock_extent_bits(tree, start, end,
9131                                          &cached_state);
9132                 }
9133
9134                 start = end + 1;
9135                 if (start < page_end)
9136                         goto again;
9137         }
9138
9139         /*
9140          * Qgroup reserved space handler
9141          * Page here will be either
9142          * 1) Already written to disk
9143          *    In this case, its reserved space is released from data rsv map
9144          *    and will be freed by delayed_ref handler finally.
9145          *    So even we call qgroup_free_data(), it won't decrease reserved
9146          *    space.
9147          * 2) Not written to disk
9148          *    This means the reserved space should be freed here. However,
9149          *    if a truncate invalidates the page (by clearing PageDirty)
9150          *    and the page is accounted for while allocating extent
9151          *    in btrfs_check_data_free_space() we let delayed_ref to
9152          *    free the entire extent.
9153          */
9154         if (PageDirty(page))
9155                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
9156         if (!inode_evicting) {
9157                 clear_extent_bit(tree, page_start, page_end,
9158                                  EXTENT_LOCKED | EXTENT_DIRTY |
9159                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
9160                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
9161                                  &cached_state, GFP_NOFS);
9162
9163                 __btrfs_releasepage(page, GFP_NOFS);
9164         }
9165
9166         ClearPageChecked(page);
9167         if (PagePrivate(page)) {
9168                 ClearPagePrivate(page);
9169                 set_page_private(page, 0);
9170                 put_page(page);
9171         }
9172 }
9173
9174 /*
9175  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
9176  * called from a page fault handler when a page is first dirtied. Hence we must
9177  * be careful to check for EOF conditions here. We set the page up correctly
9178  * for a written page which means we get ENOSPC checking when writing into
9179  * holes and correct delalloc and unwritten extent mapping on filesystems that
9180  * support these features.
9181  *
9182  * We are not allowed to take the i_mutex here so we have to play games to
9183  * protect against truncate races as the page could now be beyond EOF.  Because
9184  * vmtruncate() writes the inode size before removing pages, once we have the
9185  * page lock we can determine safely if the page is beyond EOF. If it is not
9186  * beyond EOF, then the page is guaranteed safe against truncation until we
9187  * unlock the page.
9188  */
9189 int btrfs_page_mkwrite(struct vm_fault *vmf)
9190 {
9191         struct page *page = vmf->page;
9192         struct inode *inode = file_inode(vmf->vma->vm_file);
9193         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9194         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
9195         struct btrfs_ordered_extent *ordered;
9196         struct extent_state *cached_state = NULL;
9197         struct extent_changeset *data_reserved = NULL;
9198         char *kaddr;
9199         unsigned long zero_start;
9200         loff_t size;
9201         int ret;
9202         int reserved = 0;
9203         u64 reserved_space;
9204         u64 page_start;
9205         u64 page_end;
9206         u64 end;
9207
9208         reserved_space = PAGE_SIZE;
9209
9210         sb_start_pagefault(inode->i_sb);
9211         page_start = page_offset(page);
9212         page_end = page_start + PAGE_SIZE - 1;
9213         end = page_end;
9214
9215         /*
9216          * Reserving delalloc space after obtaining the page lock can lead to
9217          * deadlock. For example, if a dirty page is locked by this function
9218          * and the call to btrfs_delalloc_reserve_space() ends up triggering
9219          * dirty page write out, then the btrfs_writepage() function could
9220          * end up waiting indefinitely to get a lock on the page currently
9221          * being processed by btrfs_page_mkwrite() function.
9222          */
9223         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
9224                                            reserved_space);
9225         if (!ret) {
9226                 ret = file_update_time(vmf->vma->vm_file);
9227                 reserved = 1;
9228         }
9229         if (ret) {
9230                 if (ret == -ENOMEM)
9231                         ret = VM_FAULT_OOM;
9232                 else /* -ENOSPC, -EIO, etc */
9233                         ret = VM_FAULT_SIGBUS;
9234                 if (reserved)
9235                         goto out;
9236                 goto out_noreserve;
9237         }
9238
9239         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
9240 again:
9241         lock_page(page);
9242         size = i_size_read(inode);
9243
9244         if ((page->mapping != inode->i_mapping) ||
9245             (page_start >= size)) {
9246                 /* page got truncated out from underneath us */
9247                 goto out_unlock;
9248         }
9249         wait_on_page_writeback(page);
9250
9251         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
9252         set_page_extent_mapped(page);
9253
9254         /*
9255          * we can't set the delalloc bits if there are pending ordered
9256          * extents.  Drop our locks and wait for them to finish
9257          */
9258         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
9259                         PAGE_SIZE);
9260         if (ordered) {
9261                 unlock_extent_cached(io_tree, page_start, page_end,
9262                                      &cached_state, GFP_NOFS);
9263                 unlock_page(page);
9264                 btrfs_start_ordered_extent(inode, ordered, 1);
9265                 btrfs_put_ordered_extent(ordered);
9266                 goto again;
9267         }
9268
9269         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9270                 reserved_space = round_up(size - page_start,
9271                                           fs_info->sectorsize);
9272                 if (reserved_space < PAGE_SIZE) {
9273                         end = page_start + reserved_space - 1;
9274                         spin_lock(&BTRFS_I(inode)->lock);
9275                         BTRFS_I(inode)->outstanding_extents++;
9276                         spin_unlock(&BTRFS_I(inode)->lock);
9277                         btrfs_delalloc_release_space(inode, data_reserved,
9278                                         page_start, PAGE_SIZE - reserved_space);
9279                 }
9280         }
9281
9282         /*
9283          * page_mkwrite gets called when the page is firstly dirtied after it's
9284          * faulted in, but write(2) could also dirty a page and set delalloc
9285          * bits, thus in this case for space account reason, we still need to
9286          * clear any delalloc bits within this page range since we have to
9287          * reserve data&meta space before lock_page() (see above comments).
9288          */
9289         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9290                           EXTENT_DIRTY | EXTENT_DELALLOC |
9291                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9292                           0, 0, &cached_state, GFP_NOFS);
9293
9294         ret = btrfs_set_extent_delalloc(inode, page_start, end,
9295                                         &cached_state, 0);
9296         if (ret) {
9297                 unlock_extent_cached(io_tree, page_start, page_end,
9298                                      &cached_state, GFP_NOFS);
9299                 ret = VM_FAULT_SIGBUS;
9300                 goto out_unlock;
9301         }
9302         ret = 0;
9303
9304         /* page is wholly or partially inside EOF */
9305         if (page_start + PAGE_SIZE > size)
9306                 zero_start = size & ~PAGE_MASK;
9307         else
9308                 zero_start = PAGE_SIZE;
9309
9310         if (zero_start != PAGE_SIZE) {
9311                 kaddr = kmap(page);
9312                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9313                 flush_dcache_page(page);
9314                 kunmap(page);
9315         }
9316         ClearPageChecked(page);
9317         set_page_dirty(page);
9318         SetPageUptodate(page);
9319
9320         BTRFS_I(inode)->last_trans = fs_info->generation;
9321         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9322         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9323
9324         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
9325
9326 out_unlock:
9327         if (!ret) {
9328                 sb_end_pagefault(inode->i_sb);
9329                 extent_changeset_free(data_reserved);
9330                 return VM_FAULT_LOCKED;
9331         }
9332         unlock_page(page);
9333 out:
9334         btrfs_delalloc_release_space(inode, data_reserved, page_start,
9335                                      reserved_space);
9336 out_noreserve:
9337         sb_end_pagefault(inode->i_sb);
9338         extent_changeset_free(data_reserved);
9339         return ret;
9340 }
9341
9342 static int btrfs_truncate(struct inode *inode)
9343 {
9344         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9345         struct btrfs_root *root = BTRFS_I(inode)->root;
9346         struct btrfs_block_rsv *rsv;
9347         int ret = 0;
9348         int err = 0;
9349         struct btrfs_trans_handle *trans;
9350         u64 mask = fs_info->sectorsize - 1;
9351         u64 min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
9352
9353         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9354                                        (u64)-1);
9355         if (ret)
9356                 return ret;
9357
9358         /*
9359          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9360          * 3 things going on here
9361          *
9362          * 1) We need to reserve space for our orphan item and the space to
9363          * delete our orphan item.  Lord knows we don't want to have a dangling
9364          * orphan item because we didn't reserve space to remove it.
9365          *
9366          * 2) We need to reserve space to update our inode.
9367          *
9368          * 3) We need to have something to cache all the space that is going to
9369          * be free'd up by the truncate operation, but also have some slack
9370          * space reserved in case it uses space during the truncate (thank you
9371          * very much snapshotting).
9372          *
9373          * And we need these to all be separate.  The fact is we can use a lot of
9374          * space doing the truncate, and we have no earthly idea how much space
9375          * we will use, so we need the truncate reservation to be separate so it
9376          * doesn't end up using space reserved for updating the inode or
9377          * removing the orphan item.  We also need to be able to stop the
9378          * transaction and start a new one, which means we need to be able to
9379          * update the inode several times, and we have no idea of knowing how
9380          * many times that will be, so we can't just reserve 1 item for the
9381          * entirety of the operation, so that has to be done separately as well.
9382          * Then there is the orphan item, which does indeed need to be held on
9383          * to for the whole operation, and we need nobody to touch this reserved
9384          * space except the orphan code.
9385          *
9386          * So that leaves us with
9387          *
9388          * 1) root->orphan_block_rsv - for the orphan deletion.
9389          * 2) rsv - for the truncate reservation, which we will steal from the
9390          * transaction reservation.
9391          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9392          * updating the inode.
9393          */
9394         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9395         if (!rsv)
9396                 return -ENOMEM;
9397         rsv->size = min_size;
9398         rsv->failfast = 1;
9399
9400         /*
9401          * 1 for the truncate slack space
9402          * 1 for updating the inode.
9403          */
9404         trans = btrfs_start_transaction(root, 2);
9405         if (IS_ERR(trans)) {
9406                 err = PTR_ERR(trans);
9407                 goto out;
9408         }
9409
9410         /* Migrate the slack space for the truncate to our reserve */
9411         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9412                                       min_size, 0);
9413         BUG_ON(ret);
9414
9415         /*
9416          * So if we truncate and then write and fsync we normally would just
9417          * write the extents that changed, which is a problem if we need to
9418          * first truncate that entire inode.  So set this flag so we write out
9419          * all of the extents in the inode to the sync log so we're completely
9420          * safe.
9421          */
9422         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9423         trans->block_rsv = rsv;
9424
9425         while (1) {
9426                 ret = btrfs_truncate_inode_items(trans, root, inode,
9427                                                  inode->i_size,
9428                                                  BTRFS_EXTENT_DATA_KEY);
9429                 if (ret != -ENOSPC && ret != -EAGAIN) {
9430                         err = ret;
9431                         break;
9432                 }
9433
9434                 trans->block_rsv = &fs_info->trans_block_rsv;
9435                 ret = btrfs_update_inode(trans, root, inode);
9436                 if (ret) {
9437                         err = ret;
9438                         break;
9439                 }
9440
9441                 btrfs_end_transaction(trans);
9442                 btrfs_btree_balance_dirty(fs_info);
9443
9444                 trans = btrfs_start_transaction(root, 2);
9445                 if (IS_ERR(trans)) {
9446                         ret = err = PTR_ERR(trans);
9447                         trans = NULL;
9448                         break;
9449                 }
9450
9451                 btrfs_block_rsv_release(fs_info, rsv, -1);
9452                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9453                                               rsv, min_size, 0);
9454                 BUG_ON(ret);    /* shouldn't happen */
9455                 trans->block_rsv = rsv;
9456         }
9457
9458         if (ret == 0 && inode->i_nlink > 0) {
9459                 trans->block_rsv = root->orphan_block_rsv;
9460                 ret = btrfs_orphan_del(trans, BTRFS_I(inode));
9461                 if (ret)
9462                         err = ret;
9463         }
9464
9465         if (trans) {
9466                 trans->block_rsv = &fs_info->trans_block_rsv;
9467                 ret = btrfs_update_inode(trans, root, inode);
9468                 if (ret && !err)
9469                         err = ret;
9470
9471                 ret = btrfs_end_transaction(trans);
9472                 btrfs_btree_balance_dirty(fs_info);
9473         }
9474 out:
9475         btrfs_free_block_rsv(fs_info, rsv);
9476
9477         if (ret && !err)
9478                 err = ret;
9479
9480         return err;
9481 }
9482
9483 /*
9484  * create a new subvolume directory/inode (helper for the ioctl).
9485  */
9486 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9487                              struct btrfs_root *new_root,
9488                              struct btrfs_root *parent_root,
9489                              u64 new_dirid)
9490 {
9491         struct inode *inode;
9492         int err;
9493         u64 index = 0;
9494
9495         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9496                                 new_dirid, new_dirid,
9497                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9498                                 &index);
9499         if (IS_ERR(inode))
9500                 return PTR_ERR(inode);
9501         inode->i_op = &btrfs_dir_inode_operations;
9502         inode->i_fop = &btrfs_dir_file_operations;
9503
9504         set_nlink(inode, 1);
9505         btrfs_i_size_write(BTRFS_I(inode), 0);
9506         unlock_new_inode(inode);
9507
9508         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9509         if (err)
9510                 btrfs_err(new_root->fs_info,
9511                           "error inheriting subvolume %llu properties: %d",
9512                           new_root->root_key.objectid, err);
9513
9514         err = btrfs_update_inode(trans, new_root, inode);
9515
9516         iput(inode);
9517         return err;
9518 }
9519
9520 struct inode *btrfs_alloc_inode(struct super_block *sb)
9521 {
9522         struct btrfs_inode *ei;
9523         struct inode *inode;
9524
9525         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9526         if (!ei)
9527                 return NULL;
9528
9529         ei->root = NULL;
9530         ei->generation = 0;
9531         ei->last_trans = 0;
9532         ei->last_sub_trans = 0;
9533         ei->logged_trans = 0;
9534         ei->delalloc_bytes = 0;
9535         ei->new_delalloc_bytes = 0;
9536         ei->defrag_bytes = 0;
9537         ei->disk_i_size = 0;
9538         ei->flags = 0;
9539         ei->csum_bytes = 0;
9540         ei->index_cnt = (u64)-1;
9541         ei->dir_index = 0;
9542         ei->last_unlink_trans = 0;
9543         ei->last_log_commit = 0;
9544         ei->delayed_iput_count = 0;
9545
9546         spin_lock_init(&ei->lock);
9547         ei->outstanding_extents = 0;
9548         ei->reserved_extents = 0;
9549
9550         ei->runtime_flags = 0;
9551         ei->prop_compress = BTRFS_COMPRESS_NONE;
9552         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9553
9554         ei->delayed_node = NULL;
9555
9556         ei->i_otime.tv_sec = 0;
9557         ei->i_otime.tv_nsec = 0;
9558
9559         inode = &ei->vfs_inode;
9560         extent_map_tree_init(&ei->extent_tree);
9561         extent_io_tree_init(&ei->io_tree, inode);
9562         extent_io_tree_init(&ei->io_failure_tree, inode);
9563         ei->io_tree.track_uptodate = 1;
9564         ei->io_failure_tree.track_uptodate = 1;
9565         atomic_set(&ei->sync_writers, 0);
9566         mutex_init(&ei->log_mutex);
9567         mutex_init(&ei->delalloc_mutex);
9568         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9569         INIT_LIST_HEAD(&ei->delalloc_inodes);
9570         INIT_LIST_HEAD(&ei->delayed_iput);
9571         RB_CLEAR_NODE(&ei->rb_node);
9572         init_rwsem(&ei->dio_sem);
9573
9574         return inode;
9575 }
9576
9577 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9578 void btrfs_test_destroy_inode(struct inode *inode)
9579 {
9580         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9581         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9582 }
9583 #endif
9584
9585 static void btrfs_i_callback(struct rcu_head *head)
9586 {
9587         struct inode *inode = container_of(head, struct inode, i_rcu);
9588         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9589 }
9590
9591 void btrfs_destroy_inode(struct inode *inode)
9592 {
9593         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9594         struct btrfs_ordered_extent *ordered;
9595         struct btrfs_root *root = BTRFS_I(inode)->root;
9596
9597         WARN_ON(!hlist_empty(&inode->i_dentry));
9598         WARN_ON(inode->i_data.nrpages);
9599         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9600         WARN_ON(BTRFS_I(inode)->reserved_extents);
9601         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9602         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9603         WARN_ON(BTRFS_I(inode)->csum_bytes);
9604         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9605
9606         /*
9607          * This can happen where we create an inode, but somebody else also
9608          * created the same inode and we need to destroy the one we already
9609          * created.
9610          */
9611         if (!root)
9612                 goto free;
9613
9614         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9615                      &BTRFS_I(inode)->runtime_flags)) {
9616                 btrfs_info(fs_info, "inode %llu still on the orphan list",
9617                            btrfs_ino(BTRFS_I(inode)));
9618                 atomic_dec(&root->orphan_inodes);
9619         }
9620
9621         while (1) {
9622                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9623                 if (!ordered)
9624                         break;
9625                 else {
9626                         btrfs_err(fs_info,
9627                                   "found ordered extent %llu %llu on inode cleanup",
9628                                   ordered->file_offset, ordered->len);
9629                         btrfs_remove_ordered_extent(inode, ordered);
9630                         btrfs_put_ordered_extent(ordered);
9631                         btrfs_put_ordered_extent(ordered);
9632                 }
9633         }
9634         btrfs_qgroup_check_reserved_leak(inode);
9635         inode_tree_del(inode);
9636         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9637 free:
9638         call_rcu(&inode->i_rcu, btrfs_i_callback);
9639 }
9640
9641 int btrfs_drop_inode(struct inode *inode)
9642 {
9643         struct btrfs_root *root = BTRFS_I(inode)->root;
9644
9645         if (root == NULL)
9646                 return 1;
9647
9648         /* the snap/subvol tree is on deleting */
9649         if (btrfs_root_refs(&root->root_item) == 0)
9650                 return 1;
9651         else
9652                 return generic_drop_inode(inode);
9653 }
9654
9655 static void init_once(void *foo)
9656 {
9657         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9658
9659         inode_init_once(&ei->vfs_inode);
9660 }
9661
9662 void btrfs_destroy_cachep(void)
9663 {
9664         /*
9665          * Make sure all delayed rcu free inodes are flushed before we
9666          * destroy cache.
9667          */
9668         rcu_barrier();
9669         kmem_cache_destroy(btrfs_inode_cachep);
9670         kmem_cache_destroy(btrfs_trans_handle_cachep);
9671         kmem_cache_destroy(btrfs_path_cachep);
9672         kmem_cache_destroy(btrfs_free_space_cachep);
9673 }
9674
9675 int btrfs_init_cachep(void)
9676 {
9677         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9678                         sizeof(struct btrfs_inode), 0,
9679                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9680                         init_once);
9681         if (!btrfs_inode_cachep)
9682                 goto fail;
9683
9684         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9685                         sizeof(struct btrfs_trans_handle), 0,
9686                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9687         if (!btrfs_trans_handle_cachep)
9688                 goto fail;
9689
9690         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9691                         sizeof(struct btrfs_path), 0,
9692                         SLAB_MEM_SPREAD, NULL);
9693         if (!btrfs_path_cachep)
9694                 goto fail;
9695
9696         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9697                         sizeof(struct btrfs_free_space), 0,
9698                         SLAB_MEM_SPREAD, NULL);
9699         if (!btrfs_free_space_cachep)
9700                 goto fail;
9701
9702         return 0;
9703 fail:
9704         btrfs_destroy_cachep();
9705         return -ENOMEM;
9706 }
9707
9708 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9709                          u32 request_mask, unsigned int flags)
9710 {
9711         u64 delalloc_bytes;
9712         struct inode *inode = d_inode(path->dentry);
9713         u32 blocksize = inode->i_sb->s_blocksize;
9714         u32 bi_flags = BTRFS_I(inode)->flags;
9715
9716         stat->result_mask |= STATX_BTIME;
9717         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9718         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9719         if (bi_flags & BTRFS_INODE_APPEND)
9720                 stat->attributes |= STATX_ATTR_APPEND;
9721         if (bi_flags & BTRFS_INODE_COMPRESS)
9722                 stat->attributes |= STATX_ATTR_COMPRESSED;
9723         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9724                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9725         if (bi_flags & BTRFS_INODE_NODUMP)
9726                 stat->attributes |= STATX_ATTR_NODUMP;
9727
9728         stat->attributes_mask |= (STATX_ATTR_APPEND |
9729                                   STATX_ATTR_COMPRESSED |
9730                                   STATX_ATTR_IMMUTABLE |
9731                                   STATX_ATTR_NODUMP);
9732
9733         generic_fillattr(inode, stat);
9734         stat->dev = BTRFS_I(inode)->root->anon_dev;
9735
9736         spin_lock(&BTRFS_I(inode)->lock);
9737         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9738         spin_unlock(&BTRFS_I(inode)->lock);
9739         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9740                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9741         return 0;
9742 }
9743
9744 static int btrfs_rename_exchange(struct inode *old_dir,
9745                               struct dentry *old_dentry,
9746                               struct inode *new_dir,
9747                               struct dentry *new_dentry)
9748 {
9749         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9750         struct btrfs_trans_handle *trans;
9751         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9752         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9753         struct inode *new_inode = new_dentry->d_inode;
9754         struct inode *old_inode = old_dentry->d_inode;
9755         struct timespec ctime = current_time(old_inode);
9756         struct dentry *parent;
9757         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9758         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9759         u64 old_idx = 0;
9760         u64 new_idx = 0;
9761         u64 root_objectid;
9762         int ret;
9763         int ret2;
9764         bool root_log_pinned = false;
9765         bool dest_log_pinned = false;
9766
9767         /* we only allow rename subvolume link between subvolumes */
9768         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9769                 return -EXDEV;
9770
9771         /* close the race window with snapshot create/destroy ioctl */
9772         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9773                 down_read(&fs_info->subvol_sem);
9774         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9775                 down_read(&fs_info->subvol_sem);
9776
9777         /*
9778          * We want to reserve the absolute worst case amount of items.  So if
9779          * both inodes are subvols and we need to unlink them then that would
9780          * require 4 item modifications, but if they are both normal inodes it
9781          * would require 5 item modifications, so we'll assume their normal
9782          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9783          * should cover the worst case number of items we'll modify.
9784          */
9785         trans = btrfs_start_transaction(root, 12);
9786         if (IS_ERR(trans)) {
9787                 ret = PTR_ERR(trans);
9788                 goto out_notrans;
9789         }
9790
9791         /*
9792          * We need to find a free sequence number both in the source and
9793          * in the destination directory for the exchange.
9794          */
9795         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9796         if (ret)
9797                 goto out_fail;
9798         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9799         if (ret)
9800                 goto out_fail;
9801
9802         BTRFS_I(old_inode)->dir_index = 0ULL;
9803         BTRFS_I(new_inode)->dir_index = 0ULL;
9804
9805         /* Reference for the source. */
9806         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9807                 /* force full log commit if subvolume involved. */
9808                 btrfs_set_log_full_commit(fs_info, trans);
9809         } else {
9810                 btrfs_pin_log_trans(root);
9811                 root_log_pinned = true;
9812                 ret = btrfs_insert_inode_ref(trans, dest,
9813                                              new_dentry->d_name.name,
9814                                              new_dentry->d_name.len,
9815                                              old_ino,
9816                                              btrfs_ino(BTRFS_I(new_dir)),
9817                                              old_idx);
9818                 if (ret)
9819                         goto out_fail;
9820         }
9821
9822         /* And now for the dest. */
9823         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9824                 /* force full log commit if subvolume involved. */
9825                 btrfs_set_log_full_commit(fs_info, trans);
9826         } else {
9827                 btrfs_pin_log_trans(dest);
9828                 dest_log_pinned = true;
9829                 ret = btrfs_insert_inode_ref(trans, root,
9830                                              old_dentry->d_name.name,
9831                                              old_dentry->d_name.len,
9832                                              new_ino,
9833                                              btrfs_ino(BTRFS_I(old_dir)),
9834                                              new_idx);
9835                 if (ret)
9836                         goto out_fail;
9837         }
9838
9839         /* Update inode version and ctime/mtime. */
9840         inode_inc_iversion(old_dir);
9841         inode_inc_iversion(new_dir);
9842         inode_inc_iversion(old_inode);
9843         inode_inc_iversion(new_inode);
9844         old_dir->i_ctime = old_dir->i_mtime = ctime;
9845         new_dir->i_ctime = new_dir->i_mtime = ctime;
9846         old_inode->i_ctime = ctime;
9847         new_inode->i_ctime = ctime;
9848
9849         if (old_dentry->d_parent != new_dentry->d_parent) {
9850                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9851                                 BTRFS_I(old_inode), 1);
9852                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9853                                 BTRFS_I(new_inode), 1);
9854         }
9855
9856         /* src is a subvolume */
9857         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9858                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9859                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9860                                           root_objectid,
9861                                           old_dentry->d_name.name,
9862                                           old_dentry->d_name.len);
9863         } else { /* src is an inode */
9864                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9865                                            BTRFS_I(old_dentry->d_inode),
9866                                            old_dentry->d_name.name,
9867                                            old_dentry->d_name.len);
9868                 if (!ret)
9869                         ret = btrfs_update_inode(trans, root, old_inode);
9870         }
9871         if (ret) {
9872                 btrfs_abort_transaction(trans, ret);
9873                 goto out_fail;
9874         }
9875
9876         /* dest is a subvolume */
9877         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9878                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9879                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9880                                           root_objectid,
9881                                           new_dentry->d_name.name,
9882                                           new_dentry->d_name.len);
9883         } else { /* dest is an inode */
9884                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9885                                            BTRFS_I(new_dentry->d_inode),
9886                                            new_dentry->d_name.name,
9887                                            new_dentry->d_name.len);
9888                 if (!ret)
9889                         ret = btrfs_update_inode(trans, dest, new_inode);
9890         }
9891         if (ret) {
9892                 btrfs_abort_transaction(trans, ret);
9893                 goto out_fail;
9894         }
9895
9896         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9897                              new_dentry->d_name.name,
9898                              new_dentry->d_name.len, 0, old_idx);
9899         if (ret) {
9900                 btrfs_abort_transaction(trans, ret);
9901                 goto out_fail;
9902         }
9903
9904         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9905                              old_dentry->d_name.name,
9906                              old_dentry->d_name.len, 0, new_idx);
9907         if (ret) {
9908                 btrfs_abort_transaction(trans, ret);
9909                 goto out_fail;
9910         }
9911
9912         if (old_inode->i_nlink == 1)
9913                 BTRFS_I(old_inode)->dir_index = old_idx;
9914         if (new_inode->i_nlink == 1)
9915                 BTRFS_I(new_inode)->dir_index = new_idx;
9916
9917         if (root_log_pinned) {
9918                 parent = new_dentry->d_parent;
9919                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9920                                 parent);
9921                 btrfs_end_log_trans(root);
9922                 root_log_pinned = false;
9923         }
9924         if (dest_log_pinned) {
9925                 parent = old_dentry->d_parent;
9926                 btrfs_log_new_name(trans, BTRFS_I(new_inode), BTRFS_I(new_dir),
9927                                 parent);
9928                 btrfs_end_log_trans(dest);
9929                 dest_log_pinned = false;
9930         }
9931 out_fail:
9932         /*
9933          * If we have pinned a log and an error happened, we unpin tasks
9934          * trying to sync the log and force them to fallback to a transaction
9935          * commit if the log currently contains any of the inodes involved in
9936          * this rename operation (to ensure we do not persist a log with an
9937          * inconsistent state for any of these inodes or leading to any
9938          * inconsistencies when replayed). If the transaction was aborted, the
9939          * abortion reason is propagated to userspace when attempting to commit
9940          * the transaction. If the log does not contain any of these inodes, we
9941          * allow the tasks to sync it.
9942          */
9943         if (ret && (root_log_pinned || dest_log_pinned)) {
9944                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9945                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9946                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9947                     (new_inode &&
9948                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9949                         btrfs_set_log_full_commit(fs_info, trans);
9950
9951                 if (root_log_pinned) {
9952                         btrfs_end_log_trans(root);
9953                         root_log_pinned = false;
9954                 }
9955                 if (dest_log_pinned) {
9956                         btrfs_end_log_trans(dest);
9957                         dest_log_pinned = false;
9958                 }
9959         }
9960         ret2 = btrfs_end_transaction(trans);
9961         ret = ret ? ret : ret2;
9962 out_notrans:
9963         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9964                 up_read(&fs_info->subvol_sem);
9965         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9966                 up_read(&fs_info->subvol_sem);
9967
9968         return ret;
9969 }
9970
9971 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9972                                      struct btrfs_root *root,
9973                                      struct inode *dir,
9974                                      struct dentry *dentry)
9975 {
9976         int ret;
9977         struct inode *inode;
9978         u64 objectid;
9979         u64 index;
9980
9981         ret = btrfs_find_free_ino(root, &objectid);
9982         if (ret)
9983                 return ret;
9984
9985         inode = btrfs_new_inode(trans, root, dir,
9986                                 dentry->d_name.name,
9987                                 dentry->d_name.len,
9988                                 btrfs_ino(BTRFS_I(dir)),
9989                                 objectid,
9990                                 S_IFCHR | WHITEOUT_MODE,
9991                                 &index);
9992
9993         if (IS_ERR(inode)) {
9994                 ret = PTR_ERR(inode);
9995                 return ret;
9996         }
9997
9998         inode->i_op = &btrfs_special_inode_operations;
9999         init_special_inode(inode, inode->i_mode,
10000                 WHITEOUT_DEV);
10001
10002         ret = btrfs_init_inode_security(trans, inode, dir,
10003                                 &dentry->d_name);
10004         if (ret)
10005                 goto out;
10006
10007         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10008                                 BTRFS_I(inode), 0, index);
10009         if (ret)
10010                 goto out;
10011
10012         ret = btrfs_update_inode(trans, root, inode);
10013 out:
10014         unlock_new_inode(inode);
10015         if (ret)
10016                 inode_dec_link_count(inode);
10017         iput(inode);
10018
10019         return ret;
10020 }
10021
10022 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
10023                            struct inode *new_dir, struct dentry *new_dentry,
10024                            unsigned int flags)
10025 {
10026         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
10027         struct btrfs_trans_handle *trans;
10028         unsigned int trans_num_items;
10029         struct btrfs_root *root = BTRFS_I(old_dir)->root;
10030         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
10031         struct inode *new_inode = d_inode(new_dentry);
10032         struct inode *old_inode = d_inode(old_dentry);
10033         u64 index = 0;
10034         u64 root_objectid;
10035         int ret;
10036         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
10037         bool log_pinned = false;
10038
10039         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
10040                 return -EPERM;
10041
10042         /* we only allow rename subvolume link between subvolumes */
10043         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
10044                 return -EXDEV;
10045
10046         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
10047             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
10048                 return -ENOTEMPTY;
10049
10050         if (S_ISDIR(old_inode->i_mode) && new_inode &&
10051             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
10052                 return -ENOTEMPTY;
10053
10054
10055         /* check for collisions, even if the  name isn't there */
10056         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
10057                              new_dentry->d_name.name,
10058                              new_dentry->d_name.len);
10059
10060         if (ret) {
10061                 if (ret == -EEXIST) {
10062                         /* we shouldn't get
10063                          * eexist without a new_inode */
10064                         if (WARN_ON(!new_inode)) {
10065                                 return ret;
10066                         }
10067                 } else {
10068                         /* maybe -EOVERFLOW */
10069                         return ret;
10070                 }
10071         }
10072         ret = 0;
10073
10074         /*
10075          * we're using rename to replace one file with another.  Start IO on it
10076          * now so  we don't add too much work to the end of the transaction
10077          */
10078         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
10079                 filemap_flush(old_inode->i_mapping);
10080
10081         /* close the racy window with snapshot create/destroy ioctl */
10082         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
10083                 down_read(&fs_info->subvol_sem);
10084         /*
10085          * We want to reserve the absolute worst case amount of items.  So if
10086          * both inodes are subvols and we need to unlink them then that would
10087          * require 4 item modifications, but if they are both normal inodes it
10088          * would require 5 item modifications, so we'll assume they are normal
10089          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
10090          * should cover the worst case number of items we'll modify.
10091          * If our rename has the whiteout flag, we need more 5 units for the
10092          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
10093          * when selinux is enabled).
10094          */
10095         trans_num_items = 11;
10096         if (flags & RENAME_WHITEOUT)
10097                 trans_num_items += 5;
10098         trans = btrfs_start_transaction(root, trans_num_items);
10099         if (IS_ERR(trans)) {
10100                 ret = PTR_ERR(trans);
10101                 goto out_notrans;
10102         }
10103
10104         if (dest != root)
10105                 btrfs_record_root_in_trans(trans, dest);
10106
10107         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
10108         if (ret)
10109                 goto out_fail;
10110
10111         BTRFS_I(old_inode)->dir_index = 0ULL;
10112         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
10113                 /* force full log commit if subvolume involved. */
10114                 btrfs_set_log_full_commit(fs_info, trans);
10115         } else {
10116                 btrfs_pin_log_trans(root);
10117                 log_pinned = true;
10118                 ret = btrfs_insert_inode_ref(trans, dest,
10119                                              new_dentry->d_name.name,
10120                                              new_dentry->d_name.len,
10121                                              old_ino,
10122                                              btrfs_ino(BTRFS_I(new_dir)), index);
10123                 if (ret)
10124                         goto out_fail;
10125         }
10126
10127         inode_inc_iversion(old_dir);
10128         inode_inc_iversion(new_dir);
10129         inode_inc_iversion(old_inode);
10130         old_dir->i_ctime = old_dir->i_mtime =
10131         new_dir->i_ctime = new_dir->i_mtime =
10132         old_inode->i_ctime = current_time(old_dir);
10133
10134         if (old_dentry->d_parent != new_dentry->d_parent)
10135                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
10136                                 BTRFS_I(old_inode), 1);
10137
10138         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
10139                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
10140                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
10141                                         old_dentry->d_name.name,
10142                                         old_dentry->d_name.len);
10143         } else {
10144                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
10145                                         BTRFS_I(d_inode(old_dentry)),
10146                                         old_dentry->d_name.name,
10147                                         old_dentry->d_name.len);
10148                 if (!ret)
10149                         ret = btrfs_update_inode(trans, root, old_inode);
10150         }
10151         if (ret) {
10152                 btrfs_abort_transaction(trans, ret);
10153                 goto out_fail;
10154         }
10155
10156         if (new_inode) {
10157                 inode_inc_iversion(new_inode);
10158                 new_inode->i_ctime = current_time(new_inode);
10159                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
10160                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
10161                         root_objectid = BTRFS_I(new_inode)->location.objectid;
10162                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
10163                                                 root_objectid,
10164                                                 new_dentry->d_name.name,
10165                                                 new_dentry->d_name.len);
10166                         BUG_ON(new_inode->i_nlink == 0);
10167                 } else {
10168                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
10169                                                  BTRFS_I(d_inode(new_dentry)),
10170                                                  new_dentry->d_name.name,
10171                                                  new_dentry->d_name.len);
10172                 }
10173                 if (!ret && new_inode->i_nlink == 0)
10174                         ret = btrfs_orphan_add(trans,
10175                                         BTRFS_I(d_inode(new_dentry)));
10176                 if (ret) {
10177                         btrfs_abort_transaction(trans, ret);
10178                         goto out_fail;
10179                 }
10180         }
10181
10182         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
10183                              new_dentry->d_name.name,
10184                              new_dentry->d_name.len, 0, index);
10185         if (ret) {
10186                 btrfs_abort_transaction(trans, ret);
10187                 goto out_fail;
10188         }
10189
10190         if (old_inode->i_nlink == 1)
10191                 BTRFS_I(old_inode)->dir_index = index;
10192
10193         if (log_pinned) {
10194                 struct dentry *parent = new_dentry->d_parent;
10195
10196                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
10197                                 parent);
10198                 btrfs_end_log_trans(root);
10199                 log_pinned = false;
10200         }
10201
10202         if (flags & RENAME_WHITEOUT) {
10203                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
10204                                                 old_dentry);
10205
10206                 if (ret) {
10207                         btrfs_abort_transaction(trans, ret);
10208                         goto out_fail;
10209                 }
10210         }
10211 out_fail:
10212         /*
10213          * If we have pinned the log and an error happened, we unpin tasks
10214          * trying to sync the log and force them to fallback to a transaction
10215          * commit if the log currently contains any of the inodes involved in
10216          * this rename operation (to ensure we do not persist a log with an
10217          * inconsistent state for any of these inodes or leading to any
10218          * inconsistencies when replayed). If the transaction was aborted, the
10219          * abortion reason is propagated to userspace when attempting to commit
10220          * the transaction. If the log does not contain any of these inodes, we
10221          * allow the tasks to sync it.
10222          */
10223         if (ret && log_pinned) {
10224                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
10225                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
10226                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
10227                     (new_inode &&
10228                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
10229                         btrfs_set_log_full_commit(fs_info, trans);
10230
10231                 btrfs_end_log_trans(root);
10232                 log_pinned = false;
10233         }
10234         btrfs_end_transaction(trans);
10235 out_notrans:
10236         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
10237                 up_read(&fs_info->subvol_sem);
10238
10239         return ret;
10240 }
10241
10242 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
10243                          struct inode *new_dir, struct dentry *new_dentry,
10244                          unsigned int flags)
10245 {
10246         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
10247                 return -EINVAL;
10248
10249         if (flags & RENAME_EXCHANGE)
10250                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
10251                                           new_dentry);
10252
10253         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
10254 }
10255
10256 static void btrfs_run_delalloc_work(struct btrfs_work *work)
10257 {
10258         struct btrfs_delalloc_work *delalloc_work;
10259         struct inode *inode;
10260
10261         delalloc_work = container_of(work, struct btrfs_delalloc_work,
10262                                      work);
10263         inode = delalloc_work->inode;
10264         filemap_flush(inode->i_mapping);
10265         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
10266                                 &BTRFS_I(inode)->runtime_flags))
10267                 filemap_flush(inode->i_mapping);
10268
10269         if (delalloc_work->delay_iput)
10270                 btrfs_add_delayed_iput(inode);
10271         else
10272                 iput(inode);
10273         complete(&delalloc_work->completion);
10274 }
10275
10276 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
10277                                                     int delay_iput)
10278 {
10279         struct btrfs_delalloc_work *work;
10280
10281         work = kmalloc(sizeof(*work), GFP_NOFS);
10282         if (!work)
10283                 return NULL;
10284
10285         init_completion(&work->completion);
10286         INIT_LIST_HEAD(&work->list);
10287         work->inode = inode;
10288         work->delay_iput = delay_iput;
10289         WARN_ON_ONCE(!inode);
10290         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
10291                         btrfs_run_delalloc_work, NULL, NULL);
10292
10293         return work;
10294 }
10295
10296 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
10297 {
10298         wait_for_completion(&work->completion);
10299         kfree(work);
10300 }
10301
10302 /*
10303  * some fairly slow code that needs optimization. This walks the list
10304  * of all the inodes with pending delalloc and forces them to disk.
10305  */
10306 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
10307                                    int nr)
10308 {
10309         struct btrfs_inode *binode;
10310         struct inode *inode;
10311         struct btrfs_delalloc_work *work, *next;
10312         struct list_head works;
10313         struct list_head splice;
10314         int ret = 0;
10315
10316         INIT_LIST_HEAD(&works);
10317         INIT_LIST_HEAD(&splice);
10318
10319         mutex_lock(&root->delalloc_mutex);
10320         spin_lock(&root->delalloc_lock);
10321         list_splice_init(&root->delalloc_inodes, &splice);
10322         while (!list_empty(&splice)) {
10323                 binode = list_entry(splice.next, struct btrfs_inode,
10324                                     delalloc_inodes);
10325
10326                 list_move_tail(&binode->delalloc_inodes,
10327                                &root->delalloc_inodes);
10328                 inode = igrab(&binode->vfs_inode);
10329                 if (!inode) {
10330                         cond_resched_lock(&root->delalloc_lock);
10331                         continue;
10332                 }
10333                 spin_unlock(&root->delalloc_lock);
10334
10335                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10336                 if (!work) {
10337                         if (delay_iput)
10338                                 btrfs_add_delayed_iput(inode);
10339                         else
10340                                 iput(inode);
10341                         ret = -ENOMEM;
10342                         goto out;
10343                 }
10344                 list_add_tail(&work->list, &works);
10345                 btrfs_queue_work(root->fs_info->flush_workers,
10346                                  &work->work);
10347                 ret++;
10348                 if (nr != -1 && ret >= nr)
10349                         goto out;
10350                 cond_resched();
10351                 spin_lock(&root->delalloc_lock);
10352         }
10353         spin_unlock(&root->delalloc_lock);
10354
10355 out:
10356         list_for_each_entry_safe(work, next, &works, list) {
10357                 list_del_init(&work->list);
10358                 btrfs_wait_and_free_delalloc_work(work);
10359         }
10360
10361         if (!list_empty_careful(&splice)) {
10362                 spin_lock(&root->delalloc_lock);
10363                 list_splice_tail(&splice, &root->delalloc_inodes);
10364                 spin_unlock(&root->delalloc_lock);
10365         }
10366         mutex_unlock(&root->delalloc_mutex);
10367         return ret;
10368 }
10369
10370 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10371 {
10372         struct btrfs_fs_info *fs_info = root->fs_info;
10373         int ret;
10374
10375         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10376                 return -EROFS;
10377
10378         ret = __start_delalloc_inodes(root, delay_iput, -1);
10379         if (ret > 0)
10380                 ret = 0;
10381         /*
10382          * the filemap_flush will queue IO into the worker threads, but
10383          * we have to make sure the IO is actually started and that
10384          * ordered extents get created before we return
10385          */
10386         atomic_inc(&fs_info->async_submit_draining);
10387         while (atomic_read(&fs_info->nr_async_submits) ||
10388                atomic_read(&fs_info->async_delalloc_pages)) {
10389                 wait_event(fs_info->async_submit_wait,
10390                            (atomic_read(&fs_info->nr_async_submits) == 0 &&
10391                             atomic_read(&fs_info->async_delalloc_pages) == 0));
10392         }
10393         atomic_dec(&fs_info->async_submit_draining);
10394         return ret;
10395 }
10396
10397 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10398                                int nr)
10399 {
10400         struct btrfs_root *root;
10401         struct list_head splice;
10402         int ret;
10403
10404         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10405                 return -EROFS;
10406
10407         INIT_LIST_HEAD(&splice);
10408
10409         mutex_lock(&fs_info->delalloc_root_mutex);
10410         spin_lock(&fs_info->delalloc_root_lock);
10411         list_splice_init(&fs_info->delalloc_roots, &splice);
10412         while (!list_empty(&splice) && nr) {
10413                 root = list_first_entry(&splice, struct btrfs_root,
10414                                         delalloc_root);
10415                 root = btrfs_grab_fs_root(root);
10416                 BUG_ON(!root);
10417                 list_move_tail(&root->delalloc_root,
10418                                &fs_info->delalloc_roots);
10419                 spin_unlock(&fs_info->delalloc_root_lock);
10420
10421                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10422                 btrfs_put_fs_root(root);
10423                 if (ret < 0)
10424                         goto out;
10425
10426                 if (nr != -1) {
10427                         nr -= ret;
10428                         WARN_ON(nr < 0);
10429                 }
10430                 spin_lock(&fs_info->delalloc_root_lock);
10431         }
10432         spin_unlock(&fs_info->delalloc_root_lock);
10433
10434         ret = 0;
10435         atomic_inc(&fs_info->async_submit_draining);
10436         while (atomic_read(&fs_info->nr_async_submits) ||
10437               atomic_read(&fs_info->async_delalloc_pages)) {
10438                 wait_event(fs_info->async_submit_wait,
10439                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
10440                     atomic_read(&fs_info->async_delalloc_pages) == 0));
10441         }
10442         atomic_dec(&fs_info->async_submit_draining);
10443 out:
10444         if (!list_empty_careful(&splice)) {
10445                 spin_lock(&fs_info->delalloc_root_lock);
10446                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10447                 spin_unlock(&fs_info->delalloc_root_lock);
10448         }
10449         mutex_unlock(&fs_info->delalloc_root_mutex);
10450         return ret;
10451 }
10452
10453 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10454                          const char *symname)
10455 {
10456         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10457         struct btrfs_trans_handle *trans;
10458         struct btrfs_root *root = BTRFS_I(dir)->root;
10459         struct btrfs_path *path;
10460         struct btrfs_key key;
10461         struct inode *inode = NULL;
10462         int err;
10463         int drop_inode = 0;
10464         u64 objectid;
10465         u64 index = 0;
10466         int name_len;
10467         int datasize;
10468         unsigned long ptr;
10469         struct btrfs_file_extent_item *ei;
10470         struct extent_buffer *leaf;
10471
10472         name_len = strlen(symname);
10473         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10474                 return -ENAMETOOLONG;
10475
10476         /*
10477          * 2 items for inode item and ref
10478          * 2 items for dir items
10479          * 1 item for updating parent inode item
10480          * 1 item for the inline extent item
10481          * 1 item for xattr if selinux is on
10482          */
10483         trans = btrfs_start_transaction(root, 7);
10484         if (IS_ERR(trans))
10485                 return PTR_ERR(trans);
10486
10487         err = btrfs_find_free_ino(root, &objectid);
10488         if (err)
10489                 goto out_unlock;
10490
10491         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10492                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10493                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10494         if (IS_ERR(inode)) {
10495                 err = PTR_ERR(inode);
10496                 goto out_unlock;
10497         }
10498
10499         /*
10500         * If the active LSM wants to access the inode during
10501         * d_instantiate it needs these. Smack checks to see
10502         * if the filesystem supports xattrs by looking at the
10503         * ops vector.
10504         */
10505         inode->i_fop = &btrfs_file_operations;
10506         inode->i_op = &btrfs_file_inode_operations;
10507         inode->i_mapping->a_ops = &btrfs_aops;
10508         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10509
10510         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10511         if (err)
10512                 goto out_unlock_inode;
10513
10514         path = btrfs_alloc_path();
10515         if (!path) {
10516                 err = -ENOMEM;
10517                 goto out_unlock_inode;
10518         }
10519         key.objectid = btrfs_ino(BTRFS_I(inode));
10520         key.offset = 0;
10521         key.type = BTRFS_EXTENT_DATA_KEY;
10522         datasize = btrfs_file_extent_calc_inline_size(name_len);
10523         err = btrfs_insert_empty_item(trans, root, path, &key,
10524                                       datasize);
10525         if (err) {
10526                 btrfs_free_path(path);
10527                 goto out_unlock_inode;
10528         }
10529         leaf = path->nodes[0];
10530         ei = btrfs_item_ptr(leaf, path->slots[0],
10531                             struct btrfs_file_extent_item);
10532         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10533         btrfs_set_file_extent_type(leaf, ei,
10534                                    BTRFS_FILE_EXTENT_INLINE);
10535         btrfs_set_file_extent_encryption(leaf, ei, 0);
10536         btrfs_set_file_extent_compression(leaf, ei, 0);
10537         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10538         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10539
10540         ptr = btrfs_file_extent_inline_start(ei);
10541         write_extent_buffer(leaf, symname, ptr, name_len);
10542         btrfs_mark_buffer_dirty(leaf);
10543         btrfs_free_path(path);
10544
10545         inode->i_op = &btrfs_symlink_inode_operations;
10546         inode_nohighmem(inode);
10547         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10548         inode_set_bytes(inode, name_len);
10549         btrfs_i_size_write(BTRFS_I(inode), name_len);
10550         err = btrfs_update_inode(trans, root, inode);
10551         /*
10552          * Last step, add directory indexes for our symlink inode. This is the
10553          * last step to avoid extra cleanup of these indexes if an error happens
10554          * elsewhere above.
10555          */
10556         if (!err)
10557                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10558                                 BTRFS_I(inode), 0, index);
10559         if (err) {
10560                 drop_inode = 1;
10561                 goto out_unlock_inode;
10562         }
10563
10564         d_instantiate_new(dentry, inode);
10565
10566 out_unlock:
10567         btrfs_end_transaction(trans);
10568         if (drop_inode) {
10569                 inode_dec_link_count(inode);
10570                 iput(inode);
10571         }
10572         btrfs_btree_balance_dirty(fs_info);
10573         return err;
10574
10575 out_unlock_inode:
10576         drop_inode = 1;
10577         unlock_new_inode(inode);
10578         goto out_unlock;
10579 }
10580
10581 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10582                                        u64 start, u64 num_bytes, u64 min_size,
10583                                        loff_t actual_len, u64 *alloc_hint,
10584                                        struct btrfs_trans_handle *trans)
10585 {
10586         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10587         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10588         struct extent_map *em;
10589         struct btrfs_root *root = BTRFS_I(inode)->root;
10590         struct btrfs_key ins;
10591         u64 cur_offset = start;
10592         u64 i_size;
10593         u64 cur_bytes;
10594         u64 last_alloc = (u64)-1;
10595         int ret = 0;
10596         bool own_trans = true;
10597         u64 end = start + num_bytes - 1;
10598
10599         if (trans)
10600                 own_trans = false;
10601         while (num_bytes > 0) {
10602                 if (own_trans) {
10603                         trans = btrfs_start_transaction(root, 3);
10604                         if (IS_ERR(trans)) {
10605                                 ret = PTR_ERR(trans);
10606                                 break;
10607                         }
10608                 }
10609
10610                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10611                 cur_bytes = max(cur_bytes, min_size);
10612                 /*
10613                  * If we are severely fragmented we could end up with really
10614                  * small allocations, so if the allocator is returning small
10615                  * chunks lets make its job easier by only searching for those
10616                  * sized chunks.
10617                  */
10618                 cur_bytes = min(cur_bytes, last_alloc);
10619                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10620                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10621                 if (ret) {
10622                         if (own_trans)
10623                                 btrfs_end_transaction(trans);
10624                         break;
10625                 }
10626                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10627
10628                 last_alloc = ins.offset;
10629                 ret = insert_reserved_file_extent(trans, inode,
10630                                                   cur_offset, ins.objectid,
10631                                                   ins.offset, ins.offset,
10632                                                   ins.offset, 0, 0, 0,
10633                                                   BTRFS_FILE_EXTENT_PREALLOC);
10634                 if (ret) {
10635                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10636                                                    ins.offset, 0);
10637                         btrfs_abort_transaction(trans, ret);
10638                         if (own_trans)
10639                                 btrfs_end_transaction(trans);
10640                         break;
10641                 }
10642
10643                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10644                                         cur_offset + ins.offset -1, 0);
10645
10646                 em = alloc_extent_map();
10647                 if (!em) {
10648                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10649                                 &BTRFS_I(inode)->runtime_flags);
10650                         goto next;
10651                 }
10652
10653                 em->start = cur_offset;
10654                 em->orig_start = cur_offset;
10655                 em->len = ins.offset;
10656                 em->block_start = ins.objectid;
10657                 em->block_len = ins.offset;
10658                 em->orig_block_len = ins.offset;
10659                 em->ram_bytes = ins.offset;
10660                 em->bdev = fs_info->fs_devices->latest_bdev;
10661                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10662                 em->generation = trans->transid;
10663
10664                 while (1) {
10665                         write_lock(&em_tree->lock);
10666                         ret = add_extent_mapping(em_tree, em, 1);
10667                         write_unlock(&em_tree->lock);
10668                         if (ret != -EEXIST)
10669                                 break;
10670                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10671                                                 cur_offset + ins.offset - 1,
10672                                                 0);
10673                 }
10674                 free_extent_map(em);
10675 next:
10676                 num_bytes -= ins.offset;
10677                 cur_offset += ins.offset;
10678                 *alloc_hint = ins.objectid + ins.offset;
10679
10680                 inode_inc_iversion(inode);
10681                 inode->i_ctime = current_time(inode);
10682                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10683                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10684                     (actual_len > inode->i_size) &&
10685                     (cur_offset > inode->i_size)) {
10686                         if (cur_offset > actual_len)
10687                                 i_size = actual_len;
10688                         else
10689                                 i_size = cur_offset;
10690                         i_size_write(inode, i_size);
10691                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10692                 }
10693
10694                 ret = btrfs_update_inode(trans, root, inode);
10695
10696                 if (ret) {
10697                         btrfs_abort_transaction(trans, ret);
10698                         if (own_trans)
10699                                 btrfs_end_transaction(trans);
10700                         break;
10701                 }
10702
10703                 if (own_trans)
10704                         btrfs_end_transaction(trans);
10705         }
10706         if (cur_offset < end)
10707                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10708                         end - cur_offset + 1);
10709         return ret;
10710 }
10711
10712 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10713                               u64 start, u64 num_bytes, u64 min_size,
10714                               loff_t actual_len, u64 *alloc_hint)
10715 {
10716         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10717                                            min_size, actual_len, alloc_hint,
10718                                            NULL);
10719 }
10720
10721 int btrfs_prealloc_file_range_trans(struct inode *inode,
10722                                     struct btrfs_trans_handle *trans, int mode,
10723                                     u64 start, u64 num_bytes, u64 min_size,
10724                                     loff_t actual_len, u64 *alloc_hint)
10725 {
10726         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10727                                            min_size, actual_len, alloc_hint, trans);
10728 }
10729
10730 static int btrfs_set_page_dirty(struct page *page)
10731 {
10732         return __set_page_dirty_nobuffers(page);
10733 }
10734
10735 static int btrfs_permission(struct inode *inode, int mask)
10736 {
10737         struct btrfs_root *root = BTRFS_I(inode)->root;
10738         umode_t mode = inode->i_mode;
10739
10740         if (mask & MAY_WRITE &&
10741             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10742                 if (btrfs_root_readonly(root))
10743                         return -EROFS;
10744                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10745                         return -EACCES;
10746         }
10747         return generic_permission(inode, mask);
10748 }
10749
10750 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10751 {
10752         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10753         struct btrfs_trans_handle *trans;
10754         struct btrfs_root *root = BTRFS_I(dir)->root;
10755         struct inode *inode = NULL;
10756         u64 objectid;
10757         u64 index;
10758         int ret = 0;
10759
10760         /*
10761          * 5 units required for adding orphan entry
10762          */
10763         trans = btrfs_start_transaction(root, 5);
10764         if (IS_ERR(trans))
10765                 return PTR_ERR(trans);
10766
10767         ret = btrfs_find_free_ino(root, &objectid);
10768         if (ret)
10769                 goto out;
10770
10771         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10772                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10773         if (IS_ERR(inode)) {
10774                 ret = PTR_ERR(inode);
10775                 inode = NULL;
10776                 goto out;
10777         }
10778
10779         inode->i_fop = &btrfs_file_operations;
10780         inode->i_op = &btrfs_file_inode_operations;
10781
10782         inode->i_mapping->a_ops = &btrfs_aops;
10783         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10784
10785         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10786         if (ret)
10787                 goto out_inode;
10788
10789         ret = btrfs_update_inode(trans, root, inode);
10790         if (ret)
10791                 goto out_inode;
10792         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10793         if (ret)
10794                 goto out_inode;
10795
10796         /*
10797          * We set number of links to 0 in btrfs_new_inode(), and here we set
10798          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10799          * through:
10800          *
10801          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10802          */
10803         set_nlink(inode, 1);
10804         unlock_new_inode(inode);
10805         d_tmpfile(dentry, inode);
10806         mark_inode_dirty(inode);
10807
10808 out:
10809         btrfs_end_transaction(trans);
10810         if (ret)
10811                 iput(inode);
10812         btrfs_balance_delayed_items(fs_info);
10813         btrfs_btree_balance_dirty(fs_info);
10814         return ret;
10815
10816 out_inode:
10817         unlock_new_inode(inode);
10818         goto out;
10819
10820 }
10821
10822 __attribute__((const))
10823 static int btrfs_readpage_io_failed_hook(struct page *page, int failed_mirror)
10824 {
10825         return -EAGAIN;
10826 }
10827
10828 static struct btrfs_fs_info *iotree_fs_info(void *private_data)
10829 {
10830         struct inode *inode = private_data;
10831         return btrfs_sb(inode->i_sb);
10832 }
10833
10834 static void btrfs_check_extent_io_range(void *private_data, const char *caller,
10835                                         u64 start, u64 end)
10836 {
10837         struct inode *inode = private_data;
10838         u64 isize;
10839
10840         isize = i_size_read(inode);
10841         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
10842                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
10843                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
10844                         caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
10845         }
10846 }
10847
10848 void btrfs_set_range_writeback(void *private_data, u64 start, u64 end)
10849 {
10850         struct inode *inode = private_data;
10851         unsigned long index = start >> PAGE_SHIFT;
10852         unsigned long end_index = end >> PAGE_SHIFT;
10853         struct page *page;
10854
10855         while (index <= end_index) {
10856                 page = find_get_page(inode->i_mapping, index);
10857                 ASSERT(page); /* Pages should be in the extent_io_tree */
10858                 set_page_writeback(page);
10859                 put_page(page);
10860                 index++;
10861         }
10862 }
10863
10864 static const struct inode_operations btrfs_dir_inode_operations = {
10865         .getattr        = btrfs_getattr,
10866         .lookup         = btrfs_lookup,
10867         .create         = btrfs_create,
10868         .unlink         = btrfs_unlink,
10869         .link           = btrfs_link,
10870         .mkdir          = btrfs_mkdir,
10871         .rmdir          = btrfs_rmdir,
10872         .rename         = btrfs_rename2,
10873         .symlink        = btrfs_symlink,
10874         .setattr        = btrfs_setattr,
10875         .mknod          = btrfs_mknod,
10876         .listxattr      = btrfs_listxattr,
10877         .permission     = btrfs_permission,
10878         .get_acl        = btrfs_get_acl,
10879         .set_acl        = btrfs_set_acl,
10880         .update_time    = btrfs_update_time,
10881         .tmpfile        = btrfs_tmpfile,
10882 };
10883 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10884         .lookup         = btrfs_lookup,
10885         .permission     = btrfs_permission,
10886         .update_time    = btrfs_update_time,
10887 };
10888
10889 static const struct file_operations btrfs_dir_file_operations = {
10890         .llseek         = generic_file_llseek,
10891         .read           = generic_read_dir,
10892         .iterate_shared = btrfs_real_readdir,
10893         .open           = btrfs_opendir,
10894         .unlocked_ioctl = btrfs_ioctl,
10895 #ifdef CONFIG_COMPAT
10896         .compat_ioctl   = btrfs_compat_ioctl,
10897 #endif
10898         .release        = btrfs_release_file,
10899         .fsync          = btrfs_sync_file,
10900 };
10901
10902 static const struct extent_io_ops btrfs_extent_io_ops = {
10903         /* mandatory callbacks */
10904         .submit_bio_hook = btrfs_submit_bio_hook,
10905         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10906         .merge_bio_hook = btrfs_merge_bio_hook,
10907         .readpage_io_failed_hook = btrfs_readpage_io_failed_hook,
10908         .tree_fs_info = iotree_fs_info,
10909         .set_range_writeback = btrfs_set_range_writeback,
10910
10911         /* optional callbacks */
10912         .fill_delalloc = run_delalloc_range,
10913         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10914         .writepage_start_hook = btrfs_writepage_start_hook,
10915         .set_bit_hook = btrfs_set_bit_hook,
10916         .clear_bit_hook = btrfs_clear_bit_hook,
10917         .merge_extent_hook = btrfs_merge_extent_hook,
10918         .split_extent_hook = btrfs_split_extent_hook,
10919         .check_extent_io_range = btrfs_check_extent_io_range,
10920 };
10921
10922 /*
10923  * btrfs doesn't support the bmap operation because swapfiles
10924  * use bmap to make a mapping of extents in the file.  They assume
10925  * these extents won't change over the life of the file and they
10926  * use the bmap result to do IO directly to the drive.
10927  *
10928  * the btrfs bmap call would return logical addresses that aren't
10929  * suitable for IO and they also will change frequently as COW
10930  * operations happen.  So, swapfile + btrfs == corruption.
10931  *
10932  * For now we're avoiding this by dropping bmap.
10933  */
10934 static const struct address_space_operations btrfs_aops = {
10935         .readpage       = btrfs_readpage,
10936         .writepage      = btrfs_writepage,
10937         .writepages     = btrfs_writepages,
10938         .readpages      = btrfs_readpages,
10939         .direct_IO      = btrfs_direct_IO,
10940         .invalidatepage = btrfs_invalidatepage,
10941         .releasepage    = btrfs_releasepage,
10942         .set_page_dirty = btrfs_set_page_dirty,
10943         .error_remove_page = generic_error_remove_page,
10944 };
10945
10946 static const struct address_space_operations btrfs_symlink_aops = {
10947         .readpage       = btrfs_readpage,
10948         .writepage      = btrfs_writepage,
10949         .invalidatepage = btrfs_invalidatepage,
10950         .releasepage    = btrfs_releasepage,
10951 };
10952
10953 static const struct inode_operations btrfs_file_inode_operations = {
10954         .getattr        = btrfs_getattr,
10955         .setattr        = btrfs_setattr,
10956         .listxattr      = btrfs_listxattr,
10957         .permission     = btrfs_permission,
10958         .fiemap         = btrfs_fiemap,
10959         .get_acl        = btrfs_get_acl,
10960         .set_acl        = btrfs_set_acl,
10961         .update_time    = btrfs_update_time,
10962 };
10963 static const struct inode_operations btrfs_special_inode_operations = {
10964         .getattr        = btrfs_getattr,
10965         .setattr        = btrfs_setattr,
10966         .permission     = btrfs_permission,
10967         .listxattr      = btrfs_listxattr,
10968         .get_acl        = btrfs_get_acl,
10969         .set_acl        = btrfs_set_acl,
10970         .update_time    = btrfs_update_time,
10971 };
10972 static const struct inode_operations btrfs_symlink_inode_operations = {
10973         .get_link       = page_get_link,
10974         .getattr        = btrfs_getattr,
10975         .setattr        = btrfs_setattr,
10976         .permission     = btrfs_permission,
10977         .listxattr      = btrfs_listxattr,
10978         .update_time    = btrfs_update_time,
10979 };
10980
10981 const struct dentry_operations btrfs_dentry_operations = {
10982         .d_delete       = btrfs_dentry_delete,
10983         .d_release      = btrfs_dentry_release,
10984 };